From b04efbd3746d4f7c2a50f0f480295dfa0495721d Mon Sep 17 00:00:00 2001 From: aminreza3303 Date: Fri, 26 Jun 2026 14:52:03 +0330 Subject: [PATCH] Add CA-TTI research manuscript package --- .gitignore | 1 + Aidaily_ca_tti_cover_letter.md | 20 + Aidaily_ca_tti_final_manuscript.docx | Bin 0 -> 162518 bytes Aidaily_ca_tti_final_manuscript.md | 343 +++++++++++++ Aidaily_ca_tti_final_manuscript.pdf | Bin 0 -> 133785 bytes Aidaily_ca_tti_final_manuscript.tex | 291 +++++++++++ Aidaily_ca_tti_manuscript_draft.md | 254 ++++++++++ Aidaily_ca_tti_manuscript_revised_stage4.md | 335 +++++++++++++ Aidaily_ca_tti_stage2_5_integrity_report.md | 83 ++++ ...ly_ca_tti_stage3_prime_rereview_package.md | 84 ++++ Aidaily_ca_tti_stage3_review_package.md | 404 +++++++++++++++ ..._ca_tti_stage4_5_final_integrity_report.md | 131 +++++ Aidaily_ca_tti_stage4_revision_package.md | 127 +++++ Aidaily_ca_tti_stage5_finalization_package.md | 47 ++ Aidaily_final_manuscript.docx | Bin 0 -> 20501 bytes Aidaily_final_manuscript.html | 450 +++++++++++++++++ Aidaily_final_manuscript.md | 434 +++++++++++++++++ Aidaily_final_manuscript.rtf | 441 +++++++++++++++++ Aidaily_final_manuscript_standalone_review.md | 160 ++++++ Aidaily_protocol_paper.md | 459 ++++++++++++++++++ Aidaily_stage3_prime_rereview_package.md | 72 +++ Aidaily_stage3_review_package.md | 213 ++++++++ Aidaily_stage4_5_final_integrity_report.md | 74 +++ Aidaily_stage4_revision_package.md | 39 ++ Aidaily_stage5_finalization_package.md | 40 ++ Aidaily_stage6_process_record.md | 125 +++++ Aidaily_v0.3_revised.md | 205 ++++++++ 27 files changed, 4832 insertions(+) create mode 100644 Aidaily_ca_tti_cover_letter.md create mode 100644 Aidaily_ca_tti_final_manuscript.docx create mode 100644 Aidaily_ca_tti_final_manuscript.md create mode 100644 Aidaily_ca_tti_final_manuscript.pdf create mode 100644 Aidaily_ca_tti_final_manuscript.tex create mode 100644 Aidaily_ca_tti_manuscript_draft.md create mode 100644 Aidaily_ca_tti_manuscript_revised_stage4.md create mode 100644 Aidaily_ca_tti_stage2_5_integrity_report.md create mode 100644 Aidaily_ca_tti_stage3_prime_rereview_package.md create mode 100644 Aidaily_ca_tti_stage3_review_package.md create mode 100644 Aidaily_ca_tti_stage4_5_final_integrity_report.md create mode 100644 Aidaily_ca_tti_stage4_revision_package.md create mode 100644 Aidaily_ca_tti_stage5_finalization_package.md create mode 100644 Aidaily_final_manuscript.docx create mode 100644 Aidaily_final_manuscript.html create mode 100644 Aidaily_final_manuscript.md create mode 100644 Aidaily_final_manuscript.rtf create mode 100644 Aidaily_final_manuscript_standalone_review.md create mode 100644 Aidaily_protocol_paper.md create mode 100644 Aidaily_stage3_prime_rereview_package.md create mode 100644 Aidaily_stage3_review_package.md create mode 100644 Aidaily_stage4_5_final_integrity_report.md create mode 100644 Aidaily_stage4_revision_package.md create mode 100644 Aidaily_stage5_finalization_package.md create mode 100644 Aidaily_stage6_process_record.md create mode 100644 Aidaily_v0.3_revised.md diff --git a/.gitignore b/.gitignore index 04fde58..0850215 100644 --- a/.gitignore +++ b/.gitignore @@ -24,3 +24,4 @@ dist-ssr *.njsproj *.sln *.sw? +.gstack/ diff --git a/Aidaily_ca_tti_cover_letter.md b/Aidaily_ca_tti_cover_letter.md new file mode 100644 index 0000000..820bf89 --- /dev/null +++ b/Aidaily_ca_tti_cover_letter.md @@ -0,0 +1,20 @@ +# Cover Letter + +June 26, 2026 + +Dear Editor, + +Please consider the manuscript entitled "Transparency Drift in Human-AI Software Teams: A Confidence-Aware Team Transparency Index" as a conceptual framework paper with synthetic stress-test evidence. + +The manuscript addresses a growing software engineering problem: AI agents can improve the visible artifact layer of software work while human review depth, shared understanding, and accountability decline. It proposes CA-TTI as an early-warning measurement framework that separates artifact score, confidence, trend state, and Human-Agent Alignment Gap rather than compressing transparency into a single score. + +The paper contributes a bounded framework, a construct definition for transparency drift, and a synthetic stress test showing how multi-signal interpretation can expose failure modes missed by a raw artifact score. It does not claim field validation; the manuscript explicitly identifies event-level data collection and field feasibility testing as future work. + +This manuscript has not been published elsewhere and is not under consideration by another journal. All content, arguments, and conclusions remain the responsibility of the author. + +AI Disclosure: This manuscript was prepared with the assistance of AI-powered academic writing tools for research framing, structure planning, draft writing, reviewer simulation, revision planning, citation verification, integrity checking, and formatting support. The author directed and reviewed the work and takes full responsibility for its accuracy and integrity. + +Sincerely, + +[Author Name] + diff --git a/Aidaily_ca_tti_final_manuscript.docx b/Aidaily_ca_tti_final_manuscript.docx new file mode 100644 index 0000000000000000000000000000000000000000..c2aafe9204712d66a74e7acb6ad3c26bcdbcfc64 GIT binary patch literal 162518 zcmeHwOKc=bdLFea$F7#kagF=Cmnn>1mAVgO@JX_ebT}3NedtSeSbVMvRFk{ zlSMW;6SLdhRau#l5r6#eS3Lj0*MIRFOZ@NafB3h5A|Lc$zl?uh=223tJ#XCe?Vtbr zxBu!FmX`SESM5Ekyz{;9zZxdNNt73HmhLTYtgbHyQQFD6aoXQo{`BZc>%nqRlwsNp zlPrz)mQSN%`TGyQ@!pFl2}?Xv4B}A{V34%fTOO3S`_bCw`ue>!OrRcP`QbN$pfPOlw7a*QKkaTT2S=wPy!+zevtBRmM31sg zHH^}7%3Ic&cVOf&??>g{@;J-8YhC@|DrUXBHgiIo^G|5@de??g8Fs@mTuqy|o7DA2zqzMTcL(7du4rwzvNJZ~ozrzrM7@KfmVQ#<}oK zm&S|EXqRJf-fB4tH+DOtdG@aXuGs(*bldHA?2w}>PXxcZowX>5xcSA}#_Gn}vVMKH zQ%3o)Mp5?x`Ng?&7?&=(xwm#Z>PLI`t0K0O7>~!R<1Lv6(0gO;51xH=D3s8OQ<6|8 zQaIl2?3QtvL=TVhFfB%54uU%k9_4Yb4B|BSuo{MGYyW9*nDxprehrSIa9HdH`@un$ z_TnyXk6Qcom!L7s)3h7CdT-6XgG7ZdMb-X1>XZtn@x7BrVVUI*2jwtnm01gTqMkB8 z@Hy<|adb+`C?;-nf6;H~Mnxx&N8}0WEx0EMi}D$GNiUAN@1LF>W8!!H&OADa$&Maw zytjry>{s<7*)28ys2c$OyK1-e4)1=pb@1ru$@0U^_04;&^?R+&dq*1^yXy~j*LPOm zxwrF2giT`%A#-zWsAqenfz2RblUzWP9?l`yUjiixA8Hb-OVecEa!Uyo8%^-}bnECZ z|L!-JmiXt}%^T(RjVn=hxGRHbf(k9#pTq6Dc{qk3877Jo55qWJ4$^Q4UjB)2>fngW zP-LOUf2_%u+X5U6cRNWgWE3@?lxOrE-Qx{)xBgz776(ZloP-JZdYpE%@lo`uTn>nC z_+f8(UH(%XUY=PCcb!4-(xkj-NN2bw@?Uc}X8;V>J>G22Fz>hR6t}l`w(spXA9o&* z*_`wEq+A#JZJIwa%VV*oWz&1}C#y(_+*_l1e&in7Y|7hxL>_c{#d$WhR zzrO!y`~I}q*oQF8Ac@oC*@LX_+}k>s0C3|G4AIMy56>Uvot^C`_cxnUX+HHH=)sbm znvRT|mKRXsF#J5rpJZv-SOc}ZG$5Pqg&olTQ9O)_;NxfX`^iMd94Hb_(a# zdKgas%uKL=VVvHU*)?av3HJas1>G9p!;9eEi<9KAJWZmHiYCa5EQz~M@N2yhie9OZ zM3(i82KZP54!2+?Q=SFz`gte=26l3FX@y2$mOw~V5(-mR%C5x~^zy+8e~%dc1VT0a_JZ{oluk5ebIA*2lUuI%HAnP zTd(;3gQ>T@)`G8rzBuGGIZ3KbrwY#`4aR%R_qKLG#+?w3iykx>d>oEI&qdl_4#K1l zZ&|0z8{&)zFiq|%Kp7dL(Bq8c#ZA*?uRJ}RNzd*=2Opb4u z`b6_3*FTc`2AOSE0qqY5-Eq)Ps{BO=8r`|S!M*Or1w6YOYU8^x41X(~YXY9oXQ%5g zu9YjKb@n!yOPGuXp;jl)QLof+6Dt9L?NZ#IYREU#j{36fu6bhOvhxTws@lBjR9|kh2yMyIsyrv zHP)`|42M>YT0$Vz;`yEF;#tOPDmc>>x*71~%lOmv8bS_O&RIqSECk;KN?7c2!E(-R zO|VSgkIRE)5&E37kbBeTcNSk}sB^Ei5VhkDcM&wWk{!e$r!gpA?-@@B>6jP96?YQg?JZY%s(4>mw1#mf&kDqnTAggTtKy&fAM?a0u0v_UGN%V`C*F;x$|{e{AahN> z8GfA-pf#hNoVSP*gl9F_Pv(jX55xSp8W8{^_#)eJf)Jm)+ztap;bP!Vo4tv!{)jPZ z1aH-D!hX1TIASB>zMZHj&O}M{FdOUV!9i~($FVP-E6zHxlg)`tpwpOP|B1y|oLU7) zbxeIthzrc+o_xX<39*iM+gbNi$=~tbSvph!ZVsCkrm}q zRc6j2FyN?u3eQ;B$O!rK+!5^fxg>(gz!+G3SFR4bvtG*U z^GV1FE++H4t6cE~(yI`9f37kyZ(>W3{ML`QkeSLCztlN?eqSUP!m4l&e*{sE>{ZE# zCEqP8&EKW(f2HY~Fn~utgP>aF>20ycFgI{T z5sK%q-7CCpl=k}&;JtO>E7QB`Uod0i_y4AbCI z9Q8AYV53*a)JH}hh+55I1(_^et@&>SN@^oY7Iuf>Xmu8<^7?$b)KqW7o3o+!+u&GX zX91@3LcBe-3M}NTFy!U?*)h^Y%VGsL zBbg?TkT@8lRsp3k_7W9Id?TLjKW+7+G|KV9ZcvTD7NTMmQj0#U0E40uMBf*{NrH%Q z$SWPp#XJ^_swBaCeq2FjVbr)Nsz_$ahgSmQ=of~w5){=C?qrz_-`|ZoQmudikm%pV zyOGmgM1ft0ylWU9JeNaM%J4;_Bm%fN8c{I{+erl386gd2fWH*v6dmIA z@c473sPjNh2%f<@6(o?cjU6vT<=(~wumcg^o8pcI(CsZk?@y$dB|eF+1-p+1F)}!j z-I@cl%RK8=@X%}h5s=1)*(#jyyegakg>OKQ5(rLQ#Mai;K3!AoBb&==A4X_oG@&VA z(UjW@3P`{lyEXHhxGy2Z8N z-e+WZ*gSHB04ySjfrYI9YT|=bc?IpT04hLJ?_@v#@H+Cxlst|VW$H_m78Q6{2?0-^ zXQ8>KVLH45ZmDb|V~ZlIzQMv;DhPvj@iCNvl7qrX2oQZt`=AYOyV4zC7i2@ArU-)w zNwDN_prQ=dhE-C=ts?Hz0<-o59J89^tGi{|y2=gnMsZx11*jz!N7U8Tapdd4euA+? zXnYWkg10{0|KRP_0Dpsn^FfgaLMXj(2oF(^^^mw{9zk6Y#ZZ<5B577Ciqg~8`Xv^O z#6Q021Pq^Cc{RWWMfkbGb*w8c>3w(YqINV0Pw4QfI2Iu>VHpKt8HJ3=7O+N7C;)_b zkSz=dL@8dC6<;ojc@Y)Z`|F{Yr>HCe`qf|t(yzfA%3SkGWFINa^X-Im1AJwqUsEQb-)K8l-R^T3rKMN zz64PfNIO&Kv|3LEh9)Hh6<&}x8{hd|x9I7NXe0zpukXttY)$|OLQ zm=x3*`;H!*r7+@a;MB{wCA7cqz-K`d`jtQ;>iYXalgZ-YSiQY5!_jtNg8DnO+i zge6$nFgqc%6&WX4pALcw8aI?MKngSoU4|0Ej2T!s5w6_Wa-0Qv&UR^EGV(ZpizNZS z$kJ<(DS8K~)SxOm6H1^?Eb`^FWiZs`VwO8Qys>7AO01Xh39=}_{DAZwU?a6^kWCJT z*nwU!6fR6GXed6|)fp-eEM>iB_YVRTzSA!=NlUbDZQVg;x0cBG$7x_YEylf@-Cka@yUqmzvgp&*D zwP=3iK;{HuLb6Mgv@$RgOtwHJYC}9hU84E)zz)%Ee1sRB7GRJ8)R<0d;3z$CLNl*r%I}NE@4YiV? zcJnTw?PG2H0kqT%L~SHUXBmTULD7914j;%cwBmr3gI$2xrKLJONqCx;`ZNceRS$uO zL_7%DUYrj(1ffn)!#qJIh>R~r6s&rBx?ue?$p%ca0xDChhAJo81sE^}h-qSpmaEvU z5Furim@QGgcK_Ylksjxr(cf`$lMQuTq;!oam#_6uSp zb`cO)2nvrh0BwD}WWg%|J5xe$ASs8)ewJJ2f;5!2e2R%KT}AJB^O z$@o<>24y5B-i|_hO&nAKpnvb5U35iz|)t90PIHG|;KHW;vD9+Y@7$TLM~ViJNmn zp5}y~&J8=c0@7NeqH4{slND#IZw2v-=bHtqqA3^$ShuY!$5RlG1E4$U6Vvay3&AoVz(gdpFkbZob>amD>kF(c42x<42AoV%o`e( z#?eWKW{v1ntI3SD2QN;4T*Wk}>9a(8M^Wj3<=|$*P9^G2y<;TJ8EN?MOwZDVwtOO zc5^30I9Wi#wFNU`Bozo^rHCm(BGRE4F~1kjzm}BKR>l4x=7#9YRSx zbxfO$do+4B?u+TO4l_r)J6z=tFN23n%#Sc7!;@eKj43Ac2=|~JB2jpa934DBw17<_ z{g)TOLlz<-=Ryo9zSf?r=?F!fG_m0C3lS;#;Jq|9 z8b2C48v@N?B4jj}7W`ZZ+|{*Bz5YP#XQr0{3z_qVkd_Mh_kgCaHu%FX zI9M3;s+{zT%ogUaNQR-kMFV&UY6T9!uB7vkN(a)l=#SP(V2oJ;%!&soMY(W%t77(o z8?1U1q9I+F60xFLMg_@E$1o`y(*Wj))x;D9_#b=l2;~0ngzz zgCCtPiDuYATxDu=w5aVfmkLPn*83EPWf?kbZ*z-$cuGW8yb)0`f}3B*c@UyUY6ZTDM#=<4fv{C9TIq>GKQduTd$>HCZr=$I(#Z@mqj93y zt%kS*ThMMN!a8*rr;`~d)opbx{&03*XR)tO=@zCollAuWV=D`~c(0#u+?H7gwsmHiSFjFBrt zMhDqr-a2zyLYh!Bpt4lO7w#_ayc>Lq9QVd0kf7okgb@TCGw*AT!z3cDQY)bnu+Ihj zoWvwa>@mru+M*|e#spwGg}um#9YI~_)`K(x`yysqb^0@FCZPoFBm%z|w6!XjmeP)G zg*QNbPIt-UZP}U&TnU^|MiIISp0LMiD@?YB)8^AbOwSxxALxz|pxOZeW}{O+ne|83 z0KWm^L`Vh1MsRukit-w)4NOWFv~jnn=t#1G0VSXqgslojK}5KK#fqs%BIQyjN`#13 zI7O1Qt2mL841J^W{!Quv1yin09sjTxeRxE|YfV=*t_nFD=Q$ifm?*L~>8fdH8Ae-x zP=mMw(vrm&j0sktJiJddFXijuyG5IE)ELPVRg<8Y#U?vG`$gysx6H4zMeSCo0=nz_ ztRf2Y(_QF!urRd8?wO*g;x74koVpJ1C~zy+Hix`d=%IXsbm=thz5phI4Ix#I|C^m6g$UxF8>Z&6N5ln%e0`ozHmYcQ_6#85VsPv&V zufn=_lLRxdFMI%fttO%nY8UFqFlmhS;A}r03+6fQ)o`4=bK~CGVGX1%xL3pC%o+fT z&pQUtN`nXUeidZi(+-xel#9#)0t~@o_ zsm>IDsvnx*$EMi^ZMI{T({N2)lTEtVF{%e(rE~Kw^8CcMt7{E>u@jiy#I4!*jC@~K zG=WWwE4r6>kF!Jtg!%?zt;f-eq!V`rkn%XHP?WHkfulcUku;1tkD9Y}9gR(%eKvbt zD%=u&buH}k@vD#pJK95kWdxN0EUP6&v=di75RTzBk{SBGrU9>!loJXe5XlkO8cb*O z2|=zmuaWGV3Om>plu{iE zWK4(6R=)3%*y}GEt?&<+@|$P2PuX{~tr?Ey}lbntE5Z1G)WRoi+@>xB{Dm{u( zpY1{J4N7lNDynf%MNtzdsfjQ4@(hFWxjF!$;f=pCPJ0al4(B<%^4uKm7>`Ua&e^ND z4Q1oLl)kyA%pR1=njE=?W0fR;?j|O46D8Q43jO5}?nbJg;ne(WyGTed+$$dE1c@{{ zddJDlI!=z5aH|rKO`bDilq|VE2a#>H%9WY5S3s1E^w>Z3h#UnRBFPyq2}M^O>d-g7 zt^)FS#D;k|7{sd}w1MW=8a$i6SvGyZyj)3Lh0-A@qev~)Jvb$OlMU;Trzk1EqM&$X z!z&xLY`Gm?4oMyaUq9D`dAtl{VrBz<7PHz@eS6wx{${~l^Ey7|G z`_FUoo8{ym)D@sUu;O(NsIg~MTcOK0?~~Sn6|}!(x;^-Dho_a2Myv2a7By3lE-jJ zMKCH&Ro~x+x8Y6|)v?OPwrW8o1=s1;PN1Du*iQj^xKwDDh7}q%lk&<8JW)z0Slh-@ zUp&_hi$Oa>(T=L@<+j^P27MkjWi?3K=*Z0ugEj*RPU^a&0bqGxdn2{>v6X0y>fCTm;Ml557UtB? zUWZSg$0WB`H-b-qpwem8z@&vMKsb;QnBDGlP+c7oixo!aEs~7Ia&+=5m_)kL%w(I8+}ucD{S` z>?65~j|b15&url+xT%(b1^Du?DJ5&?1jgqteiIV(gYgpVsdD!;WW}}D=DBeic zZvL+4<>q&jCYnlh>1x%Aky%w=vw_KMUwzk>q+oq53L{&{3ENr zt-h86bGViuqbpy4MYT$&*THybeUCiiAS032`Kr;dHW)21%4yj;9yc#nHK}%C3bCsn zn&7`vxQTlIGc6r!M0D7sLe-g=DS)uI6H6~Ds%i3RL1fRzz0dqM^YKr9Bxqi9XW^%w zgDZxig^Q0l-fYbAfH6n3tJ)24NLgRHnl(C#wdlKolM5C%&No97OZd7>8f&48X~c)c z#(h%?6?hhl@q$XbcJnY|{j64Kx;PgjJFFm4`i)o5FWu_l97Ff@43RId3Uu(v$DXbG zqOpb$yVy5cGrBWoCsY`Rf$1BYp*!<3D#|mWX9!C$e-#Hl@3Loh$OW17E~c zl8k53seplH$#Os^52C`!S9r>4E=X1_yq_+0rd;VBlVsxP{EJh-JR|phmb>j|IrNO& zC#M+JihsdfZwDhsi@~bc4F)L3Yax*U&95C5oTWL4f=B2nj;@tx1j`c&!IW+0EHV+$ ztaOkiVBpR%g%MkQd#3G~_MK+hh-rH5+iTx7*S-}@`zRj5o5BMWcI~+uV9n|%QK-=! zG#^2Z7>a_UR8HNjk0&wRTX|e3l7_B~+%kZ*CulI(FWx(H#YULkJEH0C96R*c^DCpU z!c!l%^vy+-QtmJ?g?~W5rjOTNp6MM??Vc2e=3z2m#c}|+WTqN_2phb2l1dJ$k&{s5 zfXk2vjicSYIy48znDR8KQ1{|toA|{tZ6P0@ATFjDijd8D5z7yyZ@W4Y6Pd=y3FK)# z=7bkS9fnA*Mh5b<+-@H9t&TI{kxd_FW1fw@!cHFM5P%b&f0UcUY|C-RL}i|2fa3(7 z#5s=;V)n6mJ=2!;sS%8(i1CHI@KR;y1~4L#t3dDh%QZa40w~m zNFkah@S{#L08_B(2K?-1S@8Q|d5Tw$cct<4hf&B=DBffu@1{OOVK;bIC1uU zoW6XqKSLl`!-}l64p{S>Q$eBx^9%>{ev7%Org#G78O+zU)X0W=JT{EyYNDRP9P&)u z#8D|DTw*y#$diU~;skbwXCWYAzsiYms1H!>h!jm|%H@wZ#f&NR>NK{m>dCX28LlFBuQeN9ewklZswiHu16_j^@k^K99O?4$qhzZq}SU8!BI!^o~ z2Hya!gM-V-Fs2hELm51nQMy5zcnKK0cAoKaNkdK_erI|_$|)#4&2Y`T*!yWnmB6xM zea@1u9FmYs+D5&c&ZA#u=SmO`7Y3=aRRJl1tIu6Vuhoe_FRewwx2;v~C&QDY>W1 zV?n3->}W)ISgyzesBnHXlz&uF3g5tqk~p&uRd6&Sq|8$INgKopp$N6gsms)Qlad@w zNa7)FW6Ak#>fAe3(HN&EJpF;k+fv*`%I--{$xmuUZvpIos?xss#;Wgj%-M%ir0Xr= z2XmbmdxR~UttPw$k?=NU5O)^+EK50Iu^pA;NUTE?rHT^CGqlx#d{9YrD`FaaQnt!x zmC_)TI3Q}ux!2HF6f2;*1?3b667!RmBw2>f7>?!lBUCgl0EZUBX6FK{7Nw*}w&+3{ zWs8TssUz})5bcq1(8~=Nc>uaTD4AX<018JJ4#l@|3TD|0VI9iDl8tGUM+^*+Z!4g= zUHv>8lLC7o3@8@PV-nm3Zotoe@>A5WhX^I})Oyy{8|6(azoPUg?QNQDzjOPbxg$8E zfEs$NFvnhq6WXCM_u!>yccJXV#Q;lS&SFwu8|iB!9Znz`m(E|m;5a;`oay$;7u+k7 zWXkb>Q12Ffbl$`Pe|STqre<@Yzw34R+Dhu`{14#v)ekpY{Xh}za{h+argx?GF4q^V z?+d$z-`$!@?4tWswskVaMqiv#5vcLNpZ>V)>kH__TYmV6d3~`zTMwZcpPi~Wko*lk zbFnGoaI89}qc1ihJN30je8*m%5lH_;T+I?DWdo#6SZ2BBp*|l^;SK-36=-^(&ibN* zqRgmn!&p{jk5@)k8Nh}}Df}<0TwkmfEf79OOPBdIr^HFx87o!fY*g2$TlsV=MZONI zD6$6rH1Wk=o*^X;+-ihw)Gn_~by|U*h<122d(t1no#A%*gPbnHQ3Ss`bi%?j2yZC9 z4KqY=V|>*`kqOFWY~`c&WxLMOrb|@6&fAU(5d0f>=-nGP>)!Y@)ffsiHgQ%RnnC21 zb7`rX1T=SuQ4h?(EOHce6sFGH^ZeSY^bJCO^OomxM#;kD%9tqwSQW>s#!+t}Cq8<$ zqox<7tWHmpeIW0J)b}#%J|1Z@qczO{rX#wHbVTZR@R(7CO%IV`>l)vbotJ z9Y>3ydcvd+s3ZfX{2*oO(v>b>td@y^gG4K`G*$blyn^Yyc`;-AVTtEqOzA3c7bA z`&7}I;que1V7mJ#hwm`GLVUukJ`*LGy`^khaYfZ~+C5l*}^Sz2;SG>(HTN2FSMJNnEDUp>_zZ=mg) zkJU9YD@>;wHr$eCqroskgEBN>(nu^C1>(-)CWbRPM|v^}uAnV@)!3`X4tC!P+(vE% za20BBFJp=j@{Ao+7VVh5_+SneRlm>fL1??-UB%v2T;uVrKx3a~taLoD3g2l}_$aTC z%*Jk9Mc9kb@~Mmb%^W>JJ3fca>kc)~y(O+<-Lu_v!G#l8`h?>^LN;C>6{j5>XO;D# z{w5lsMJXP~3Htf=rL%aMG5oCkZg4BhaV%a@_Q-zQko^el_@D%%*9-d4F_9j9ev1+P z7HA@@`f+-_*z3hL+TIGi_`2Q2g7W^<6LC^aoO%MEkZxyEoTg~#iOOB4v59RdY7@W% z^bSqfB3ZhI+RN+>oQ64?grcIM!D0BtW+VXH>}T=oHd{a7=NZ|MgRuV+8Dg)RnYxP^ z8#$Bcikc)mHAFmMEqj0gLGg->QgiUnuFABmaDL`!l;w0csA_)QIy1@`JHJ=&2Eops9EwhshPzy}jEJ4-()h1U&c#ZEB?{GTdnB^9!dMu^VRvRd5oc zR|Q(^Tl{45afky~BJ=~t*fGuv%Y%i!;klTM3+@i9VM@ z_5z#rm}7wZnlr|_b2c45kE#~}c0*^>I7kK8e1w1jp91#F7%mFHncMfWEMt4<#M!fK z9IpOSxcW=%AJ42(N3*;%ExvT4@s~fYKnm=3t%&9&PT9hy+%?u>z{T^o@@P(+@r|k{ z8zT|1XE#OKTt*5Zi3zIL`8leBKJDa8Cs&LQpdX+DcduskM0}f6GSTB69YjI!RAB z-WPkX*><_fbrv(6N6ni>&F57<$`;VeKv3HvvXW;m&lOxbhfQDfyJzEPGZa0o+$^me zBoV50+c*x7@g&d8yOc|NBC=&e7n+F0^@;n55~n| zo$x3}@B)KH<48n0>dk%X;A-WuTtFuf&!=n5@GcOzHg8sUev(w|uGfg$dJWoZ&<&!z zID(~BBl^gu_`Dp=Q{M>Z+Up%@RFxw{?$xRH2=^(hWe8AusUu|o4U#1t(dgHfU-Kpt ze!|`Agc*EDl5kXbMcda?!X?|pS%@JU&$ak^2Cv16dh?X=5z2TB4RoTQkHeX5c-GhM zc)h!30WfWPw8UztnvKml&;0{ZlMsLf^PDlB&StQBvu)&Cgrj~hXQ1f zGmn|_llucRNO!pHWmVpid_J>RE#y!$Co&l0z}RfO8teyY)JIYzs$I-IL6KxQqZ1Je z$-_jByB<~E^={CMn6q9E5Q^gq8i|$2{O*NSA~VU5>6u0q5-}5Z0G05GG*5jNXoD5< zi-c>bzItojg>=xe z=&Mpy-wt^5#)JCi4cub@Z?jY1zEc^f2erb?zI>>^jN8OH950uR({Y%0nZ%yXT$3WA zdV{(B4etytya|#*$=$rUv|l##ulEo*Y0aTM*EafY@D_1*ef8cBu-Vef-9p`_)Y2su zB2d_3@+->wIw6znb#A?4f7{$+BZ>7dn&QzUPz zWcxtZh=ALX$$6BCg6|~dyTFZ|2P|E-dmMbHU%tDC>)V^W?qUTW(7()5yF^1!A_CIj z+S%22Sz-#kneww>wg`@Y~!qFE|0n4u$in!Dw7h=ZD&sfy7A#}k`s^JjDgkJywP@M+* zcs}az1zx_u%fUA>v+i<%*T((L2X1v66oE$Nz|HbwUf~E&gFIa&y|Q4+dc18pb!B&op-V*hMt_jW9Xb^C!U4gEDP1{Xpf|G zwA`?q&m-y0BIz^MzG_VUCd1uldZ*RGc8A7wDAG;oquD@rlZT=I>T}CVC{i734N|<1B2c? zR<3)j^t`ghNzW@?ScogdHwg0$HdKR+tySIo_R4Zx2C!i>x_Rvvp+Fx)g9=+*t_z>{ z18WvWw3_Y*_8kZ=)!)(k*>2X)<}D^~$4{zb-eNKkA$#|_sn<=FZt8iZ=aq&t$iH9@ zZ=P3ftnR4OLcGP4m1QiZoFH@WZ_5XFMPdpv^PEP{GJzL)>EPyTzX z^jPV!(jPx_-Wg-wf>#!Nx)ypX$y-SirEyTDodOZ^QeJVp4JrP5E6jc07`2~7S2#e- z8%^G5+FSPL(9KifD6(>zW~F{;a>zTyMK^p(o=?eht)^Qc+r268O?ek7Z%0zSH=4%& zG>J3)MDCQh_kGG2b^fw9pS=0x&8N;>LE0~``E=i#Pu_fTHC1muInV$weAei)k2zgW z%t@q*KPtrJO4Ub&JWZwhaXHQceMU=A;52e{IL*)vS2slty0JRlh3%79^x=xS&j>G} z7iyg1;Nk!}~XVtS*P<;Y1j0iJbx>4ymhW~*QW4>>Qv42Cmp+%qKA zA(U(eT;^HYpR*aP4V=*)bREKTNd)4u^Aw(G*61~>|H)UU7hCP4A#yvx!7gZ3+!;ij zWAw}h`~@GRX}0v5tlsJ(lD25oTZr{8jNCoh# zanwZ2c{B-n-vnihzTeWKS5{1UF~|W(9Cbo8 z6%S&ZE(+}6VMqA6l}CM?mlNUDf*H7ju+eSDi_;XJOF*X^mf@V(;adg!R>8hiaQ7ff zkd;10LF7VP4Qoxnw+c48Hw1+DPu;A4>WS{hYLp?tNw=LUl4*QLR)e6DS>#dlhEcSF zwxv~Ws#6|CeTS$9LB9DKzK=68A3S#Yu6i|Q^3z;aw0eYW&9SBR#fC*)Uu&F$XfMyO zu7RD}I`-VtC#WcF(SP3x{4{MzTUiRv%tDx`=m^#qMaPrx7F@H5X(^%nm2mRyGJN1e z1wIT}2EJRc?-qQvdH0*wy3uRU+k)N})T)eIXMID%z5b zo+_XNak-xkqqJ=G!%0-w&aoKUd;hClz;H(#5x;^(`HTbl^-aU&8q|mEP zdK8V4?3B9{{0KV~m1WdWrujv`O{SqTD7iM1OUci!|L{haq_gLWK~`ZG+Y!nuSR{Z7 zlAw&jVJktwOoHl%u%DvZ1vLUvS;9hzmM)YCdU-gEfCa}WbSVQp5-ShPXeg~vm=~wr z_$2ODz@sk5%|=l!1w84oybYg$RT0px3d<t|mB&%e@aSMaEu1S()qv2(1nSRlrR)oWOkfu|Ros=b zmLZrbcxRYme^BpJ263-ux%KtVX%G)lopk~}sVcTc#c5{%;}6wAz_cRlMdWwje?=+S zIH-Vkt+3BAEkg$Sc6d}Tc<3$J4i~C3{ZY@smO;ZpewJ(H(T}Sb;!M`wufn{$2pmhR zp}VnyBXM%VY6*}xTa-C7mS6G2)2{is$>YI7C&s%`9{C~?5wBS#E6PBbHE2XTMUg3ti#B-@8_4XNW{*?{E{^{rr3>#%3+eB+!cD|00PEUYm~JS z#?T%-P#ZwqPgwK~bp|3k9-&YfHM<##piyERp$3^IWDe3+YD&(+b|}T*0z+70(0)DL zPdDQBPck$LfKmsYE^qb14z2TUSPa@3bsuFmgDGSzqpl4WaNFUiHZwe~HgNStRDy)X z$pA!zs?1Vy?G!0cdrt=l%F&~U&Jl7|$G8kD_!vAq8}L*N)i#>|?BbMay0or9a|4>q zG_ebae@d!F@XnjFZ1=_1fHbv6$cc8H1hCd&dBGW`I}|`JpnC(51{VKn@CZgFQ7I@! z_)tQ#IfR;4T10Fg zu>y0usGrR8dHV0BnLiS{5cO|cNw1%gc^a1pcM3QV7d zf@~Z1nAn#s7@6_}I(1aTA=;bJN`-k#ms&I;OCQH2JT+$FdP>HVLeCzNf*K8}ktT_9 zD~50#z(E3b7O;tmS%q=|8aaYB6(gV$y$s5v$pUKRW!yUjgxS4BJ7R=m()6JNM|Y!8 zvZyhJ-W83xyZDW}MY|gw90;mfJ=fF5KxlsmX@Kjh-uPQ6To~FMuSJeUJZV}uO zo0HII)(Pc9a@}plc>8IXr~LBoWM6&lKB85Qd$`t zipZ`NWuU1>OyynScyxs4Y_LL~EofuEb#(NU`;bQv0C3JJmm?82kHG2RwR1+q7j#y^ zRHcEA5SCHOhukhqlVpjV$DA}`fuEFv6$e$Uip1_QlCPx6nAi}u;{?9#AcLdGV3C*r zO*KnVks!hO2yC-5-}XVm4R(SXt3K;H+n_% z(5R;gS2lDMf^+@sCqG3v#dHJ$1QaX70DW}RNg6*%JDrfYRc45O3WLy&Y9f2co2h|| z-H2!Rp9V?~&oqNnd`OU7bs!>6f+|DY7s62;NeIFE(&Yup_lTj2tXG0Wl-KDFvL4+v ziD4_%cu3$635g|j5iCNu3b4-rR|K%23Xq>b5bMVv6FSGJ|AP<*EQB8yh!SMU37*<- zzz_R_v2ao-A6I>^c7@3?YDS&22#!&yijCCtRGC-sZ z!NM?|`aWSNqY?T?i8lru1&?$h&0)TOmk7%^IlEnG1FB6FEqe{t_9=kc(;c{Ai^3%0 zZA}P4G0UL(gFqSShgi_1IRs+2VtC7JwKSW=A1EP=90;p{ z6lM4dB2{#FQ6`B?Luzk~f<=^0ykhyn0%Xw5>G?!4kO&UjFfjQP8jD;*8vVw*J0{dWn zqZm<3#aDngMx+UyC7=TYoEE|SZY>@ode7=8{J z!H#@-L3fH|pTyAF5rLt8W%EnKnM3tMDeF{k`Y}vO=@%G)%3&G6U~Wia8;=(|K+AO? zwP`0+NN+@Av7VEBa;Ly|2ty(5D)=D_Mk3LoGnS_#v8y3FK8MAc<)=Ee0tPx25p=4Q z=A}qNrRGv4HuEIG*l0o`XW5P`VC$a*C zvdS50Lr??5sB7AuGF*qoQ*Gga0f%J;^9igfM1Xtfll zvU7vPNwve8i;>A-IANPQw+W`(lwiF5k^Yw5kBEUrX zl>Vc1rD)?J{sMHO{WoWjNZBZm5m0lnpSW`t!&G@0b)E*ZG#cGuaA8Ka#USjz1lqnj zy@s)2*}&k0Yy}onsm`CnG9+bAb*>e}6IHHdbUN-lf^waKT7Nr7I0>|YJ9Vy_rVfbz zr0Oz1Pkc=p5gVA?joRRFVOg7D;jkU|D@k;#QFIMML`W=BE2&DD%x&{`bsQ7T&p z)ZgowG(+Bntpee;L{=2YlmjK>=NHH#1Y5DbV3pTpU4hWFY;q*rtOXSB6Z&lsz)4Pv zSbKE9o{l3<6G)-<{!Yn_CvatGfWs+8PC(aVn|?OH8gN4d#YhRw8R8U>Ams88ZVfP@rfJ1+`@C5d<7amaVcCZD8Z+Tb9IBhJ)duM&1J6QD?XkX1lQ( zJW*cc&#>k@#RAl{xm8B-&^DQpOnF3d24*b6Ah|hx*-gkZsJ z`mQx6t*;qG2Nwf1*uaIO#ff0IdLB$&iY!pE;%kz+O~wih3SJ8|9WRh4=>{#R=X4T; zpXDlV6=jK8v)UF`taAbMa&v$;iZLWHU|N#3Ao@-QhuBVnp5?K2$7|*te8{{@&B~zw z8zhX8;7N_wXu+;QLBM@pb=jU>yY~pZ+Kf>;r+FPbc;HpwOVo-7kU`~uW(-_}P;#N7 zh;dFvr<#vxl;PDCl`5o@LZt^1qp>_hpmq%@wg>eVBGmOC8hs#mN~c7N+65(65}SC0 zR%GA+Qw|4sLGBXc&`NGz!6mQz;VZ1|5!3Z1i#JDUa{Y{wsg4C<-E3ixiy*(OclIXS z0RN)s5N4Z=gC7XP$6=93gMS6p1lEeVk*YU6!E%?>X;3DiGk#j~HX@L@Rmgi74ckP% zokg^RH8udeKz`@ak#9_+bt8dwDlVzI{--dR5*J4G&Nm@`f4C_`@ zF4-u+(Qzno%`W_F;44j9g{zXGMu}HKARua=4IvLqg$;~qASvy6b4UZU*>GGt?!*HQ zn?}1%qBN`E2{*xiOi2ebh?A~NGA~f#1vp7>i}((Sbs!gv*wqP()8nj4--6N(m;JOV zBzK+K(w~xiE4lCcRR4l~Q;ZaR5RMjr-Zv+DjzG7FRKq`2ivnedC@!K+P#u_ZWsx}F zA-U;&2i$ECTj2_J!1zj3BRt_BJ$UUYP+-(oNG;P|ZF6)7{X8@&#{q-k)W#!t1l0b9 zOibhU_7I@Qb~`7{oH?O}VpOzn1}GZmqtM?Yr=O@-!Qui_Q-_~mB|tJ1%OICf1u)$M zPtNx^U_xYqn4Uss2u^e8h{|sP6BPeEV{1$o(Pd{Yo|!5jm0eO)y&e)92`ovoRk$`u zrFRO`v>ejXjAB8t4u~5~rCFwmtKa}@)KEFLloRVgJry`im8)Dto0LNpf-FUF$t2kb zE23&eA2d|hag-!2mF0*q%EUkrpl<8~GZmhwEy$XqKowH@7(Qqp3OjJj*mdVFQvMGxSVN6}zmRmD(qZKfZn9aF=f)gRM zmU?yd=4Sc-Bx{0{d-_3PiqsQySV7AMQ6SEx+?a&b&7guX5@WnKDoq65 z9nnN2D0B7wxn&sm1WXA4zxujv4)4%h-Ih)F)b#1h)7)F9LDMnd;Dk8yu8TRoIsS2g zND}fYAl6VpI08Jp*ol1AvPmvWk25P)m4L)04LDmjV&Rhw8e>Hf9Z-;q& z5`tt-!j<3=zJDhij@}KPL`eQy34XwPg10u;H@Dta2mm$`nNXUUI{if z-)U{!-$(Qpdo#iMa8Fj|<}3ZXDe z)Obud6~xgKL>q<_OyFdl;pwoQox`ejK!6(n&&K+Fu&hmdyYCRv zB&)KT;Tcv1^AEKE^s+jSzYZO0hGp0j7|n*hbcPPiJyYrMf%DYs(%}(YRRz&1c#*;T za2|3&N&~%ifL27VtxF;os&4Tb-SP)8eI0;3CYw6Lxb8pLxi-dSczrRU7@R!%H%ou; zPnYD+H;rOYuti<5dNr*;4YD9XzF6%-3q0sbkD^{kFQnF~L?P7I)FqLO4J{E{J=qy- z689jtkr`0#7jX#9n@R^_ZCU;yyM~b>qY{a&I}zTZjAgn~6LY2Kq|_N<{zXdE=4* zZgvwuB=|+p6*o48A<(<9x4gM+zq~+aI83Abh*NA%#E!{{G~dcA6Pt>e2_j;(ObOpx zCR;>`xa#93KRtz+@u4*lEGAG}G!UzQuIb$ZL{=*8W6kYkW6&mSJ1|V5GQX)HqWR5% zEI-P8GkB~kXjY{u0kN)_Hcw%Y@*_{#y9Y@jVV~p-Xh1Wh}(>eEI?ao$d3VX}j z53HUoLm?hPjQQx~Z0|LV6m}PzyVP#H$CA?a1G5YnzWAahr241DY}MxZTV-h)D07Wl zRiy*e6@E5|(oca7^^RTN7H-`W+SX)$;eOUAtBCyd_H*mr%>Ar?TFicKoxh(pK5@PK zS-^YyVcbQ!OTCvjzsH-{RxpNn_qfLO%k1&yY?`fqTFf4ApTEbh-FuBRdrS6M$!NXD z_~LTZvwXS6|4Z$$1??HWylH=0%pUKYzsEafonHw$R!Z5td)(LwEAs)U>9Th*`(>QM zc+E7G#|nV2#D1;2eD^|{ta15LG}$D1S62es81aLV_3m@KwOVUGP8y;A^U+WL{O^AY z-f8@?-f6UnMWQl1x>vMA?Kq1yMNeQ$TO+d$pQca!N+Zk?k zEf|ypDv=&*Y#*zm%1=PECqw&Nn|02mAqTkuqa8WM)&KqX|9$f3-$wvmS?EdBYd&n* zsO04X6{|xo6QXu?dRAVdEjpBjWJp8?A|V{hbvyRkRMM7*nev%U42s(?IjH}=MugLidr7>uk{pgHJ99=IEObN`NO z&^z~EQx}j2Zsp!k#dOS`dCVfC$-zu;mG8@$MaVRWop_T_>x5nOuQyg}LQpTgV|LNv zZI*e!ws|vPyH2C!Jir+=d_4v~58$?L25{GDQJe?3%W3=b0B-w+0C$b*{&|4AoGLyK z;C5~XaM!7ep9i?h>B;i|?%vG+?m89a^8j}_C3YUb-M<;YU8fd%9^fvgVa@}%2R8(` zYg92E!13c*^$*@+D69B;!VB))S)XF4o!%S?PiO_QF@llQ(A}nvpu_%l1pPt9dM2_H z|Kor7m(RDqwzR}Qzh<)&O`E+d<|pc578@42Gx40)j=mHKhD-!srj&N5K-Nd>Zfeq8Dyakb>VlNQi?C zkI}1RT49lJEO*#|Tn|-dsFFF71ksCe{KE33H84cQ&oC&;Pa|Nuy+o*JoS!z2%|*^b z0`I7%GLul>FPH)9jur^Ig+qn)M4XEkKxAi;bFo!-^-S!g3+8ce zk@K+ig!Md#`g0dKn#Z1@?cv_?<5x(jMn~3jcJ%%tXD?~;W(s@D@^wxA-Vh*alvk5` z%D#kpa__CVR5zK4^7TJ`_3!@tmzS3K=bwSA{W01+u!{U8DWdqd9QxACOoq556Vt`) zt&-DF=?psCJ|N^#SCFM(W8k&RJhF{bRx})rcoeLZ(661O-HB?Ao*6Wy?LsGEhn{C1XZh0WQ0D|ZWUzVO+LR{?UA-Opf_VUEp6A)9Sd#;j&Zpxk zm|*x4pKbveVK9_vJAA=Qg$j+vp~S!yz1?_2s`qu4i}`S#C${@Z``3rkDNprUz=)>lN?%s*= nF}a9I6|)}c_eNivrf@$0!7mBrerxGlOMe3r`9CM1ilzS#CMu>T literal 0 HcmV?d00001 diff --git a/Aidaily_ca_tti_final_manuscript.md b/Aidaily_ca_tti_final_manuscript.md new file mode 100644 index 0000000..7962351 --- /dev/null +++ b/Aidaily_ca_tti_final_manuscript.md @@ -0,0 +1,343 @@ +# Transparency Drift in Human-AI Software Teams: A Confidence-Aware Team Transparency Index + +## Abstract + +**Background:** Software teams increasingly coordinate work through human communication, project-management artifacts, code repositories, code reviews, and AI-generated updates. Coding agents and AI assistants can generate pull requests, issue comments, documentation, summaries, and decision records. These artifacts can make a project appear more complete and traceable even when human review, shared understanding, and accountability are weakening. + +**Objective:** This paper introduces CA-TTI, a confidence-aware Team Transparency Index for conceptualizing and flagging possible transparency drift in human-AI software teams. Transparency drift is defined as a growing divergence between visible project artifacts and the human team's shared understanding, review depth, and ability to explain or control the work being produced. + +**Methods:** We revise an earlier Team Transparency Index based on coverage, consistency, consensus, timeliness, and completeness. CA-TTI preserves these artifact-oriented components but changes the output from a single score into a multi-signal framework: artifact score, confidence, trend state, and Human-Agent Alignment Gap (HAG). HAG is treated as a family of alignment indicators, not as a single validated latent variable. We evaluate initial measurement behavior using a synthetic stress test across five controlled scenarios. + +**Results:** In the synthetic prototype, the raw TTI rule did not flag the designed warning cases in artifact drift and fluent hallucination scenarios. The CA-TTI prototype warning rules flagged all such trials, with mean warning lead times of 3.0 steps for artifact drift and 1.5 steps for fluent hallucination. In a noisy-interaction scenario without designed artifact failure, raw TTI warned in 7 of 8 trials, while CA-TTI produced no warnings. Ablation checks on the same synthetic observations indicated that removing trend logic substantially reduced warning lead time in the two designed warning scenarios. + +**Conclusion:** CA-TTI is proposed as an early-warning measurement framework, not as a validated productivity metric or validated field index. Its central claim is that transparency in human-AI software teams should be interpreted as a multi-signal condition. Artifact completeness can improve while shared human understanding declines. Future work should implement event-level datasets, test inter-rater reliability, run sensitivity analyses, compare external baselines, and validate the framework in real teams. + +**Keywords:** human-AI software teams; team transparency; AI agents; software traceability; Agile software development; shared mental models; measurement framework; transparency drift + +## 1. Introduction + +Software teams do not coordinate only through formal records. They rely on stand-up meetings, chat threads, issue trackers, pull requests, code reviews, release notes, documentation, and informal memory. These sources create a distributed picture of work. A decision may begin in a meeting, be clarified in chat, appear indirectly in a pull request, and never be reflected in the issue tracker. This fragmentation creates a persistent gap between what the team informally knows and what the project system formally records. + +Earlier versions of the Team Transparency Index (TTI) were designed to measure whether communication, project artifacts, and confirmed team knowledge were aligned. The original use case was an AI mediator for Agile teams: a conversational system would ingest stand-ups, chat, issue data, and Git metadata; extract candidate decisions and action items; detect mismatches; and request role-aware confirmation before writing back to the project record. In that framing, TTI was mainly an outcome measure for evaluating whether the mediator improved transparency. + +The rise of coding agents changes the measurement problem. AI systems no longer only summarize or remind. They can generate code, open pull requests, update issues, draft documentation, propose decisions, and produce fluent explanations. Evidence from AI pair-programming and agentic code-review research suggests that AI systems are becoming part of everyday software workflows, although their contributions, adoption patterns, and review needs differ from human work (Peng et al., 2023; Zhong et al., 2026). This can improve visible traceability, but it can also create a new kind of transparency failure. The artifact layer may become more complete while the team loses human understanding, review depth, accountability, or trust. + +This paper calls that proposed failure mode **transparency drift**: a gradual divergence between visible project artifacts and the shared understanding of the human team. Transparency drift matters because artifact fluency can look like coordination. A project may show more comments, cleaner issue records, faster summaries, and more complete decision logs while team members are less able to explain why decisions were made, who endorsed them, whether AI-generated changes were deeply reviewed, or how errors should be corrected. + +The central contribution of this paper is CA-TTI, a confidence-aware measurement framework for human-AI software teams. The framework is not a validated field index yet. It is a conceptual and methodological proposal supported by synthetic stress-test evidence. The key shift is: + +```text +CA-TTI = artifact score + confidence + trend state + Human-Agent Alignment Gap +``` + +CA-TTI is not intended to replace human judgment or rank individuals. It is an early-warning framework for team-level inquiry. Its purpose is to show when an apparently orderly system may be drifting away from shared human understanding. + +## 2. Related Work and Positioning + +CA-TTI sits between four research streams: Agile coordination, software traceability, human-AI teaming, and governance of workplace monitoring. + +Agile software development depends on frequent communication and shared context. Daily stand-up meetings can support awareness, coordination, and monitoring, but prior work also shows that their value depends on team context and meeting quality (Stray et al., 2016, 2017, 2020). Agile transparency therefore cannot be inferred from the presence of ceremonies alone. + +Traceability research raises a parallel issue. Software traceability is valuable, but it is often created ad hoc and after the fact, which limits its practical benefit (Cleland-Huang et al., 2014). Socio-technical congruence research similarly links coordination needs to actual coordination patterns in software teams (Cataldo et al., 2008). These literatures show that software transparency is not only a documentation problem. It is a relationship between work dependencies, communication, records, and shared understanding. + +Code review adds a further signal. Code review is not merely a gate for defect detection. It also supports knowledge transfer, maintainability, and shared standards; developers judge review quality through factors such as feedback usefulness, clarity, and reviewer expertise (Kononenko et al., 2016). When AI agents participate in code review, the question is not only whether feedback is syntactically correct, but whether humans still understand, contest, and integrate that feedback in meaningful ways. + +Human-AI teaming research provides the theoretical bridge. Human-AI teams require more than a tool-user relationship; they require coordination around shared goals, roles, communication, trust, and team cognition (Berretta et al., 2023). Shared mental models are especially relevant because a team can coordinate effectively only when members maintain compatible expectations about tasks, roles, and system behavior (Andrews et al., 2023). Empirical work on human-agent teams also suggests that communication, explicitly shared goals, trust, and perceived team cognition shape performance and collaboration (Schelble et al., 2022). + +AI-supported project work adds another layer. Recent work on requirements engineering, cognitive agents, and Agile project-management support suggests that AI and machine-learning systems can extract structured requirements information and simulate or support Agile project-management roles (Cinkusz et al., 2025; Umar et al., 2025). Human-centered AI research warns that useful automation must preserve human control, safety, and trust (Shneiderman, 2020). Workplace surveillance research adds a governance concern: measurement systems can become instruments of monitoring and pressure if they expose individual behavior without appropriate safeguards (Ball, 2021). + +These strands suggest a measurement problem rather than only a tool-building problem. The question is not simply whether AI can produce more complete project artifacts. The harder question is whether a team remains genuinely transparent when AI participates in producing those artifacts. + +## 3. Construct Boundary: What CA-TTI Measures + +This paper uses transparency in a team-level, socio-technical sense. It does not equate transparency with explainability of an AI model or with the number of project artifacts. Four forms of transparency are relevant: + +| Transparency form | Question | Example signal | +| --- | --- | --- | +| Artifact transparency | Are project records linked, current, and complete? | Issue, pull request, commit, and decision-record links | +| Process transparency | Is it clear how work moved from discussion to decision to implementation? | Trace from meeting or chat decision to implementation artifact | +| Epistemic transparency | Can affected humans explain and challenge the work? | Human review depth, confirmation, correction, and shared explanation | +| Governance transparency | Are provenance, access, consent, and accountability clear? | AI provenance labels, appeal paths, access rules | + +CA-TTI is designed for the intersection of these forms. The original artifact-oriented TTI mostly measured artifact transparency. CA-TTI keeps that core but adds confidence, trend, and HAG so that artifact improvement is not mistaken for full team transparency. + +The intended use is a team-level diagnostic and audit protocol. It can feed a dashboard, but the dashboard is not the primary contribution. The primary contribution is an interpretation rule: CA-TTI warnings should trigger team inquiry, not sanctions or individual performance assessment. A team uses the index to ask: "Do our artifacts still reflect what we jointly understand, review, and control?" + +## 4. From Artifact Transparency to Confidence-Aware Transparency + +### 4.1 Original Artifact-Oriented TTI: Artifact Transparency + +The original TTI used five components: + +```text +TTI = 0.25*COV + 0.25*CON + 0.20*CSN + 0.15*TML + 0.15*CMP +``` + +| Component | Meaning | +| --- | --- | +| COV | Coverage: eligible communication-mentioned tasks or decisions linked to an issue, pull request, Git artifact, or decision record. | +| CON | Consistency: status, ownership, blocker, and decision claims match structured records or are explicitly reconciled. | +| CSN | Consensus: high-impact decisions meet predefined role-confirmation thresholds before writeback. | +| TML | Timeliness: documented updates occur soon enough after the relevant event. | +| CMP | Completeness: action items and decisions include required metadata such as who, what, and when. | + +This score is useful because it makes artifact transparency operational. However, it remains vulnerable to two problems. First, the score can become overconfident under sparse or uneven data. If only one source is available, a clean-looking score may reflect missing evidence rather than real alignment. Second, the score can reward documentation hygiene even when human alignment is weakening. AI agents may improve links, summaries, and completeness while reducing review depth or obscuring responsibility. + +### 4.2 CA-TTI: Multi-Signal Transparency + +CA-TTI keeps the five artifact components but changes the output format. Instead of returning a single score, it returns four signals: + +```text +artifact_score +confidence +trend_state +human_agent_alignment_gap +``` + +The artifact score answers whether visible project records are linked, consistent, confirmed, timely, and complete. Confidence answers how much evidence supports that score. Trend state answers whether the system is improving, stable, deteriorating, noisy, or insufficiently observed. The Human-Agent Alignment Gap answers whether AI-generated artifacts remain aligned with human understanding, review, endorsement, and control. + +This separation matters because the same artifact score can have different meanings. A high artifact score with high confidence and low HAG is a strong signal. A high artifact score with low confidence is fragile. A high artifact score with rising HAG is the core risk case: artifacts look clean while shared understanding deteriorates. + +### 4.3 Signal Flow + +```text +team events + -> communication, issues, pull requests, reviews, decisions, AI actions + -> artifact score / confidence / trend / HAG + -> warning state + -> team inquiry and governance response +``` + +CA-TTI therefore should not be read as a scalar ranking. It is a bundle of diagnostic signals. The warning state is useful only if the team investigates the underlying components. + +## 5. Human-Agent Alignment Gap + +The Human-Agent Alignment Gap (HAG) is the distance between what agents produce or record and what the human team has actually reviewed, endorsed, understood, or accepted. + +HAG is not an anti-AI measure. It does not treat agent participation as harmful by default. It asks whether AI contribution is growing faster than the team's capacity for review, confirmation, and shared understanding. + +HAG should be treated as a family of subdimensions: + +| HAG subdimension | Definition | Observable indicator | Current prototype coverage | +| --- | --- | --- | --- | +| Confirmation debt | AI-created or AI-modified claims lack explicit human confirmation. | Unconfirmed decision records, auto-updated tickets, unendorsed summaries | Partial | +| Review-depth gap | AI-generated work receives shallower review than comparable human work. | Review rounds, comment depth, test discussion, reviewer expertise | Not yet implemented | +| Attribution ambiguity | The team cannot tell whether a claim came from a human, an agent, or a mixed process. | Missing provenance label, unclear author chain | Not yet implemented | +| Explanation gap | Affected humans cannot explain why a decision or change was made. | Post-hoc explanation checks, reviewer challenge outcomes | Not yet implemented | +| Correction gap | AI-generated artifacts require more correction, revert, or clarification. | Reverts, follow-up corrections, rejected suggestions | Partial | +| Trust and safety divergence | Artifacts improve while trust, psychological safety, or willingness to challenge declines. | Team survey or retrospective signal | Not yet implemented | + +The current synthetic prototype does not implement full HAG. It implements a narrower `hag_proxy`, a hallucination-alignment proxy derived from unsupported claims, confidence-artifact mismatch, and missing evidence. This is a useful first stress-test component, but it should not be interpreted as a complete operationalization of human-agent alignment. + +## 6. Synthetic Stress-Test Design + +The present draft uses a synthetic stress test to inspect whether CA-TTI behaves sensibly under controlled, author-designed warning conditions. This is not a field validation study. It is a measurement-behavior test. + +The prototype compared a raw TTI-like score against a CA-TTI score. The generator produced 8 trials and 10 steps for each of five scenarios, yielding 400 trajectory-level observations. Each row included: + +```text +scenario +trial_id +step +raw_tti +artifact_score +confidence +unsupported_claim_rate +evidence_coverage +actual_quality +failure_label +``` + +The scenarios were: + +| Scenario | Purpose | +| --- | --- | +| Clean baseline | Raw TTI and artifact-centered signals should agree. | +| Artifact drift | Designed case where artifact quality deteriorates before interaction quality visibly fails. | +| Fluent hallucination | Output remains fluent and confident while artifact quality and evidence coverage collapse. | +| Low-confidence good artifacts | Useful artifacts exist but system confidence is muted. | +| Noisy interaction with stable artifacts | Interaction-level noise should not be treated as artifact failure. | + +The prototype scorer computed trend from a four-step artifact-history window. A negative artifact slope reduced the trend score; fewer than four observations defaulted to a neutral trend score. The prototype `hag_proxy` combined unsupported-claim rate, confidence-artifact mismatch, and evidence-coverage gap: + +```text +hag_proxy = + 0.52 * unsupported_claim_rate ++ 0.30 * max(0, confidence - artifact_score) ++ 0.18 * max(0, 1 - evidence_coverage) +``` + +Confidence was calibrated downward when `hag_proxy` rose: + +```text +calibrated_confidence = confidence * (1 - 0.65 * hag_proxy) +``` + +The prototype CA-TTI score was: + +```text +ca_tti_score = + 0.45 * artifact_score ++ 0.20 * calibrated_confidence ++ 0.20 * trend_score ++ 0.15 * (1 - hag_proxy) +``` + +Warnings were generated when: + +```text +ca_tti_score < 0.58 +or hag_proxy >= 0.42 +or trend_score <= 0.42 +``` + +For transparency, the raw baseline used the same numeric warning threshold for the raw TTI-like score: + +```text +raw_tti < 0.58 +``` + +These weights and thresholds were chosen as prototype stress-test settings, not as validated field cutoffs. A deployed system would require calibration by domain, team workflow, and evidence availability. The present results should therefore be read as behavior of one specified prototype configuration, not as evidence that the selected weights or cutoffs are optimal. + +## 7. Synthetic Results + +The results are summarized below. + +| Scenario | Designed Warning Trials | Raw Warn Trials | CA-TTI Warn Trials | Mean CA-TTI Lead | +| --- | ---: | ---: | ---: | ---: | +| Artifact drift | 8/8 | 0/8 | 8/8 | 3.0 steps | +| Fluent hallucination | 8/8 | 0/8 | 8/8 | 1.5 steps | +| Clean baseline | 0/8 | 0/8 | 0/8 | n/a | +| Low-confidence good artifacts | 0/8 | 1/8 | 0/8 | n/a | +| Noisy interaction with stable artifacts | 0/8 | 7/8 | 0/8 | n/a | + +In the two designed warning scenarios, raw TTI did not warn, while CA-TTI warned in every trial under the prototype rules. In the noisy but stable scenario, raw TTI produced warnings in most trials, while CA-TTI produced none. In the low-confidence good-artifact scenario, CA-TTI did not treat low confidence alone as failure when artifacts and evidence remained strong. + +An ablation check using the same scored observations suggests that trend logic contributed substantially to early warning in this prototype. Removing HAG did not change warning counts in this small synthetic test, but removing trend reduced lead time below the designed warning point in both warning scenarios. The check is preliminary because it was not pre-registered and uses the same synthetic data. + +| Designed warning scenario | Full CA-TTI lead | No HAG lead | No trend lead | Artifact-only lead | +| --- | ---: | ---: | ---: | ---: | +| Artifact drift | 3.0 | 3.0 | -0.25 | 1.0 | +| Fluent hallucination | 1.5 | 1.5 | -0.88 | 0.25 | + +These results support only a bounded claim: separating artifact score, confidence calibration, trend, and alignment-gap signals can make author-designed synthetic warning modes visible under one prototype configuration. They do not establish construct validity, external validity, comparative superiority, or real-world predictive accuracy. + +## 8. Use, Misuse, and Deployment Vignette + +CA-TTI should be used as a team-level diagnostic and early-warning framework. It should not be used to score individual developers, rank teams, or evaluate personal performance. Such use would distort behavior and create surveillance risk. + +The governance rule is: + +```text +A team should not be considered more transparent if artifact transparency improves while psychological safety, trust, or human-agent alignment declines. +``` + +Practical deployment requires safeguards: + +1. Report CA-TTI at the team level, not the individual level. +2. Separate artifact score from confidence, trend, and HAG. +3. Preserve provenance for AI-created or AI-modified records. +4. Require explicit human confirmation for high-impact decisions. +5. Treat psychological safety and workload as hard governance constraints. +6. Allow participants to challenge, correct, or reverse AI-mediated records. +7. Avoid manager-facing dashboards that expose individual prompt behavior. +8. Retain raw event data only as long as needed for team-level audit. +9. Treat warnings as inquiry triggers, not sanctions. + +A safe deployment vignette illustrates the intended use. During a sprint review, a team sees that artifact score has risen from 0.71 to 0.84 because more pull requests, issue links, and decision records are present. Confidence is moderate, but HAG is rising because several agent-generated summaries were accepted without review and developers cannot explain the rationale behind two decisions. The correct response is not to identify a low-performing developer. The correct response is a team inquiry: review which AI-generated records require confirmation, add provenance labels, ask affected developers to explain or revise the decision records, and adjust review policy for future agent-generated changes. + +This use case also shows how gaming should be handled. A team could try to increase confirmations without improving understanding. CA-TTI therefore should not reward confirmation count alone. Confirmation must be linked to review depth, provenance, correction rights, and the ability of affected humans to challenge the record. + +## 9. Discussion + +The shift from an artifact-oriented index to CA-TTI changes the paper's contribution. The original contribution was a protocol for evaluating a conversational AI mediator. The revised contribution is a measurement framework for human-AI software teams. + +This change makes the paper more durable. Tool designs will change quickly as AI agents evolve. A measurement problem will remain: how can teams know whether AI participation is improving shared transparency or merely improving the appearance of traceability? + +CA-TTI addresses that problem by refusing to compress transparency into one score. A high artifact score with low confidence should not be interpreted like a high artifact score with rich evidence. A high artifact score with rising HAG should not be interpreted as success. A temporary noisy transition should not be treated like confirmed transparency decline. These distinctions are the main value of the framework. + +The framework also clarifies how future validation should proceed. A field study should not only ask whether CA-TTI increases during an intervention. It should ask whether CA-TTI predicts cases where teams report lower alignment, lower trust, lower psychological safety, or weaker shared mental models despite improved artifact completeness. + +## 10. Limitations + +This paper has six major limitations. + +First, the synthetic data are not real Agile data. They test measurement behavior under designed scenarios but cannot establish construct validity. + +Second, the current prototype is trajectory-level rather than event-level. A fuller version should generate or collect event-level data with fields such as event type, source, actor type, linked artifact, confirmation status, delay, metadata completeness, review depth, autonomy level, correction events, trust signal, psychological safety signal, and missingness flag. + +Third, HAG is only partially operationalized in the current prototype. The implemented `hag_proxy` focuses on unsupported claims and confidence-evidence mismatch. The broader construct should include review depth, confirmation debt, attribution ambiguity, correction/revert rate, explanation gap, trust decline, and psychological safety decline. + +Fourth, the thresholds used in this prototype are not validated field cutoffs. They are stress-test settings. Real deployment would require calibration, sensitivity analysis, and stakeholder review. + +Fifth, the prototype has not yet been compared against independent external baselines. The raw TTI comparison is useful for showing how the revised design differs from the earlier artifact-oriented score, but it is not sufficient for a strong comparative claim. + +Sixth, the literature base is stronger than in the initial draft but still selective. A full journal submission should include a broader review of AI coding agents, team cognition measurement, software repository mining, and responsible workplace AI governance. + +## 11. Future Work + +Future work should proceed in three stages. + +First, improve the synthetic generator so it produces event-level software-team records rather than only trajectories. This would make the metric easier to explain and closer to the eventual field setting. + +Second, conduct coder-based plausibility review. Human reviewers should inspect generated scenarios and judge whether the synthetic events plausibly represent healthy teams, documentation-only improvement, human-agent drift, noisy transition, and true transparency decline. Inter-rater reliability should be reported. + +Third, run a small field feasibility study. The goal should not be causal proof. The goal should be to test whether CA-TTI can be computed reliably, whether HAG can be coded consistently, and whether team members find the outputs meaningful rather than intrusive. This study should include independent baselines and sensitivity analysis for the proposed weights and warning thresholds. + +## 12. Conclusion + +Human-AI software teams need a way to detect transparency drift before it becomes visible as project failure. Raw artifact scores are not enough because AI agents can improve the visible record while weakening shared human understanding. CA-TTI responds by separating artifact transparency, confidence, trend, and Human-Agent Alignment Gap. + +The first synthetic stress test supports the plausibility of this framing. Under the prototype warning rules, CA-TTI flagged designed artifact drift and fluent hallucination scenarios that raw TTI did not flag, while avoiding warnings in noisy but stable conditions. Ablation checks suggest that trend logic is important for early warning in the current prototype. These findings are preliminary, but they show why a single transparency score may be insufficient for human-AI software teams. + +CA-TTI should therefore be understood as an early-warning measurement framework. Its purpose is not to declare that a team is productive or well-managed. Its purpose is to ask whether the team still understands, reviews, confirms, and controls the work that humans and agents are producing together. + +## Declarations + +### Data Availability + +The synthetic stress-test outputs used in this manuscript are available in the accompanying project experiment output directory. The current manuscript reports only synthetic measurement-behavior results and does not include human-subject data. + +### Funding + +No external funding is declared for this manuscript draft. + +### Conflicts of Interest + +The author declares no conflicts of interest for this manuscript draft. + +### AI Disclosure + +This paper was prepared with the assistance of AI-powered academic writing tools. The AI pipeline included research framing, structure planning, draft writing, reviewer simulation, revision planning, citation verification, integrity checking, and formatting support. All content, arguments, and conclusions were directed and reviewed by the author. The author takes full responsibility for the accuracy and integrity of this work. + +## References + +Andrews, R. W., Lilly, J. M., Srivastava, D., & Feigh, K. M. (2023). The role of shared mental models in human-AI teams: A theoretical review. *Theoretical Issues in Ergonomics Science, 24*(2), 129-175. https://doi.org/10.1080/1463922X.2022.2061080 + +Ball, K. (2021). *Electronic monitoring and surveillance in the workplace: Literature review and policy recommendations*. Publications Office of the European Union. https://doi.org/10.2760/5137 + +Berretta, S., Tausch, A., Ontrup, G., Gilles, B., Peifer, C., & Kluge, A. (2023). Defining human-AI teaming the human-centered way: A scoping review and network analysis. *Frontiers in Artificial Intelligence, 6*, Article 1250725. https://doi.org/10.3389/frai.2023.1250725 + +Cataldo, M., Herbsleb, J. D., & Carley, K. M. (2008). Socio-technical congruence. In *Proceedings of the Second ACM-IEEE International Symposium on Empirical Software Engineering and Measurement* (pp. 2-11). ACM. https://doi.org/10.1145/1414004.1414008 + +Cinkusz, K., Chudziak, J. A., & Niewiadomska-Szynkiewicz, E. (2025). Cognitive agents powered by large language models for agile software project management. *Electronics, 14*(1), Article 87. https://doi.org/10.3390/electronics14010087 + +Cleland-Huang, J., Gotel, O. C. Z., Huffman Hayes, J., Mäder, P., & Zisman, A. (2014). Software traceability: Trends and future directions. In *Future of Software Engineering Proceedings* (pp. 55-69). ACM. https://doi.org/10.1145/2593882.2593891 + +Kononenko, O., Baysal, O., & Godfrey, M. W. (2016). Code review quality. In *Proceedings of the 38th International Conference on Software Engineering* (pp. 1028-1038). ACM. https://doi.org/10.1145/2884781.2884840 + +Peng, S., Kalliamvakou, E., Cihon, P., & Demirer, M. (2023). *The impact of AI on developer productivity: Evidence from GitHub Copilot*. arXiv. https://arxiv.org/abs/2302.06590 + +Schelble, B. G., Flathmann, C., McNeese, N. J., Freeman, G., & Mallick, R. (2022). Let's think together! Assessing shared mental models, performance, and trust in human-agent teams. *Proceedings of the ACM on Human-Computer Interaction, 6*(GROUP), Article 13, 1-29. https://doi.org/10.1145/3492832 + +Shneiderman, B. (2020). Human-centered artificial intelligence: Reliable, safe & trustworthy. *International Journal of Human-Computer Interaction, 36*(6), 495-504. https://doi.org/10.1080/10447318.2020.1741118 + +Stray, V., Moe, N. B., & Bergersen, G. R. (2017). Are daily stand-up meetings valuable? A survey of developers in software teams. In H. Baumeister, H. Lichter, & M. Riebisch (Eds.), *Agile Processes in Software Engineering and Extreme Programming* (pp. 274-281). Springer. https://doi.org/10.1007/978-3-319-57633-6_20 + +Stray, V., Moe, N. B., & Sjoberg, D. I. K. (2020). Daily stand-up meetings: Start breaking the rules. *IEEE Software, 37*(3), 70-77. https://doi.org/10.1109/MS.2018.2875988 + +Stray, V., Sjøberg, D. I. K., & Dybå, T. (2016). The daily stand-up meeting: A grounded theory study. *Journal of Systems and Software, 114*, 101-124. https://doi.org/10.1016/j.jss.2016.01.004 + +Umar, M. A., Lano, K., & Abubakar, A. K. (2025). Automated requirements engineering framework for agile model-driven development. *Frontiers in Computer Science, 7*, Article 1537100. https://doi.org/10.3389/fcomp.2025.1537100 + +Zhong, S., Noei, S., Zou, Y., & Adams, B. (2026). *Human-AI synergy in agentic code review*. arXiv. https://arxiv.org/abs/2603.15911 diff --git a/Aidaily_ca_tti_final_manuscript.pdf b/Aidaily_ca_tti_final_manuscript.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2b3d13c2c5fad59d782692a477eb537a21fc8c42 GIT binary patch literal 133785 zcma&NQ+Fj3EY z*Z(m`wrk25d6C$))iSF!HjhFMk9&#`mR8fZo}-BaVe)g&yyKVNM(_AdF}BVwx37zK z_Xh};!^9uSprf1KJh>_L%8^(#d}q0^>Zt>;k!2% z4&r^V{o#pTEZ&MjPCU;ERLZ`7Cw5cpnTs+}{UJmQfUZfcWsdd9nJC(#)G7#0JX)^c)$i#&^5<-$(J2 zA*O&y!XNvqW3U0rAuIvoAZnt0@B)D%yL-X|M&Nc-X^9>ir@^A8^ zOKKu-u7xBX&1naO_gxP14%{tA38c+SfemubWR`EZDp~D9kLcmBV{~@gr2S{^BW<{? zdRm+T*}Smy@uF0M77P&2)PAhL28@E&F@7cnlF;eA4F}w2#ll)}arh=yQdX@YV;2I> zI#}yy)g_Hh97e`a3|SxiWRR$vbQvPvn~?FDZ@r!kyqZ=EU^p>3C^{NuN`<16 z263@~n>Pjh0~ubsX8~&A+yS3MAHW zP5NN2ujvdnseVS|hmLvzs z_Zcr5G623Q)3P;Y=a}bq@zspcrsAwkq4dK`nf|n&)dJF%^gbT&+B89^p?VZybq?Dm zEATHFEqiG9CF|!NbMV18RW_yj?o3=6;T3#&SAiT_Rd{npT@%B?@_X7e? zd5i~RVr%^W%g%qV|5vD3{zr<%#sTjR1?N!9m0?a8}@MlzFedd}<+^7qXSR8iz^yjaU84+fNC0 zw^GjFdP%2J=?Kt3*>nJ}ux6>)e1HrjPEgSb$U9Jaj<4&T@{#$PaI7uGAT)nLzSkc7 ztNg$>EvkjoT>j9`N)9kuEHpb+Tla~mr`=TW)k}vb_kTSaWm%GYs(WkIg#cv;@>rfl z&Q?#r`15Hjwp?Q|*#H`))c+)5e?H@{rj>y4+)>Ddc&$z2RGgsAqeQMw$}$a2>g9=YB60&U+s*Z!a5$jbOXS)eQ|jI96b zJoac!#Im#@_nuNcfh?W$0tF%f9{_b=zuO5SB$^&Y`P@;}5z?!S5NGw6;|=$PC^StF zYfP!ZR!k9MdLWuYDOgsgRDmW7xZ`?qvY7YT-lCOo4`-N5!)?eb0UH#(9up_Hr=+xEur7)p914esyaf7 z0?87lbZB^x^5VdzNz=OeN%>8y_6TmL_e{cXN2qy8b2T%dx7FsqQ+pw?Fii0UFK`Tv zgDrc*u`Y4Y0>8U?&!~A@2lY%)JsOzvk0Q-r;o0nR!CjGd095WKyU@DXs5NI&+Be(tTOId3a2*;ZopYWdZb6@fu9 zkXHrv@Kdn+Z?~k^WYYIj)AfW_6D7Kfb$RmMd^t7qd-X1hk~d1%XQh3{=A#?tCB^H+ z9K=orsc9O2<%?V4O9RBHx5`Hc@9SD`s~OfT>+&V7RE)Yai1?#SD#8#OR}xfZV&ATq zISUlFG0^7Kvx9^-I2|}~UHEsPyWeCl@+tlAH3gv|6ky~?Dq5v$9$3JEFiSY0aga`(N+S%iX z-rSF&9(Z#L?y#robAnOB_0oeQ1cjcy0+z@fD-?z9vVNu_2Z1y>c>*5mm$wzJC`B0F zTDHl&A|xeNQ}U0&D&7>H<8dO6Ife~lyD;Mnn;i&jJqWCS88h{rofe}ea!rWwNtAh! z)^wMqVBV3tp@NpF?Xc z*+42cTmZCBD~}-Yl`MClHABfY>Q30J_Z83;4F>6mZnmrDLgcVj@56AxFx^EebXvB= zAThB!uJg^!zo=uG#ebOY%x6k~M-^Ud+iVt?&xNnO{YV2~qgX(lqAB{4--rDp;Y6S! zjFn<^{9wp1lgz^4ZxnJyZxEkIXC7VThT~@#oOnVUTk42U1TdjyfnE(%wr24F5VnZI zJm_xrK-fl8Gan8FN-knMxUm^Hy(KY}_Gw`woKpA0BW>jdGWb>G29Om)LPqld!!PU1k{U zc1GiEKEbVgPwp@Q{~0wIRrEKkVAV(!l%;<9t z91i%k(pjV}SGM?Ge+pLvF&&qk+hBPD_l}xRoQNX=#K={P~z3_jH>Uo7_fM z8BCw8dGm{^7>S@Q5@yZUd0GO|;Rwp)ql9Cu>_4eeT}Q`LH{!v0s$|#b<|E+2{f9G8 zN5WMDKVf(%HF%(YDpG_;*@v0nsp^1wvxOrcC{lCytsTd= zYq7jogpt&}LBIkk0g*9ACRF&@E*B8}<%SF1OY^2B?*g*&#SvV^9NchC7S#8?ZR<(8 zmSN7JN1*8$NA)Z%>E9FKlzX&#Kj z7XqRNsQVbkrs(hkuI%LeF;d-)0=$J$7lPKO-LBr{G%~<&IUFdT{#hkumE=)OL*^{q z`cj0L4iG_dB*m)CU`pQr23+X{s@b4=UcgW2>s%!8E-~tD5;_q2Dvit$YwRngd&fdu zl#6p1^q)Dj@ua&G`w=a25WB;A*}jAEQ25@^L3E6_NL}Lfo9syd)=Z&f+H@u2^L=D1 zw!2}hnU+C0smhFBWAzdu=HS7mS!xP-C76&V9(~7hPe?+=jA;2>#?sWPlZ&T>-Pog4 zqe!*1(Rou$0o)(gFGKN$oLXmOZB$IyD*UNv69!}LNmZVcA;awBeXLy;;mscV>2VQm zn$7o&LR^*0`rSM}aROA7MxI^5d`yAngfLw1gGME?V(DfR)dp-=LSnRY!NCYjo4z8mq3zDggyHMGdT@#W&2~~| zC}aq(0_HrMUi08|Is~-o~^2c^V;gWX_^F3pxCy_g3=SXA1`BlU$EXJ{X zFly-DQWr9C4;J$L6#i6|9dy&Axks5oe%Jg@{tItbl~rdO3?mt+F9QSUWEUGmu|i01 z2RnudGMC7d>~vRt1dY&fsMQX%+z`8zyw@KRI!!mKbYSGRJT@@bFBC|(qHWby`BBW( zD%XmR_6078{l!^1Q=@0&GNbbnVdgv8YmgU0Kk0n%rYXDc6qFJ4vlsBx)0o1uIr zKVpH*V(~+|o;4&tVqKe@_~I@-=4LPCO*ax7H3xzz+wDKx;7qMmT$?m?RT~l^uuP8p z(+zgJ?9CKTb_*bMseCi9f;616@#J5FM)*g+lb#u&Ze=>vJh1UYlBr-mVDu2YXsScJ zJD^NMakP}|mhwYWU4f%~oF(Y4Z?|}-jFVt5VqpNrrG>b>4eBs2o*1vxQ#qH{?X!%r zn@Q)e3dTxVI=b%&;zT?tu3kBN4|hYfo%4ST#1m~*8Qy5v&NNvOz&OMV)iU39G%GT* z(*6%u_NB32pR6I~A2-o1#2oB(=xc_xf!r_3)o@?+^K-WL5_15_wl-aLM;B073An1P z@Jm>jnr5jYV%4q)z!-ThX(Kn|bL^M8nm@K$s(sRN8>FF6N(BL^o9-vuvgfLpY|Uid zbLG_y!a{4+EB{O37HZm@Nwdz)*sXto$SNPM*`N|;ZesdKnN`)b8#_#$I8PWG9vauf zZBaDbwc`-7;h-TR5-NaeQ?WNe@3#NrXS$+cpXb8K(&5L~ z#e*j*(V|~n+IoY!M=rFkf@nX#_bET%qd#AU{})N zpx0;LK+#=L@nGZh<5z*S-rLvf#u!bLw`HUW%L$SyFV|o!qb*WsA`;$7D`~q7ABC;I zPh^==aPOIlu|E%QHeN56FYh!yzCB;(X;Xy1g$$ZHO_S6mQw($09T94{McckUzo#-( zxct?-Kc02TPMdl?K7%g}PhC86*Bvx-goL}%Q=4$87EcX64==p7B)Z!hsCDmY?_0j; zaFuoS-*wY3+dNp47+axCHhJ8&4D$CR#acz>v6VvCV|~8wn&oft6cV6$cm|dA^4PDtweph+fSHWS-EJf>5wHxwzwG8iVdx7uicQMosgF%A4VnuVc&J^#Tjw5qx6b=E zOfs=F^o?{|4jNqLycJ50MoDq5T+hOw*$qzhinoJ1r8ir5Mc}0Rj1bmYl6bguHpeA0J^|stHuX>qgYzE$j^WF+-tgn8qdw*qnHN49#pQ@6j5*@s(H-1Ll+g&KQTz)x-F* zKZKFAH5Gvn@9iDJYDr9VpK%FGM%6=U;vV`1NJADb9WILnH?F#iPfigMsv6mpw#m8q z``{M<2NJ=^QNd%RHe4DHuyexo)0(u7r8j={m(a<_7#meXM%t#o2={L7&{Ni#Vwhv` zI^;RbXU3Bz_p%02?%{bxC4>g#nMlDRon^YOuLxO|-krgzSs971diSkTqoWY+WyFMF zQ=dl8`*h#3;QgVw+MduFf`26FT5dQI_D87&LV7LV&wr!mG4^npORs3k_4B5e3wt0Y|NX4@^m`iti%R4|PZuKa) ztFH6ZrBu*Hk%-UZuMYs4D(}F5&5SuNLt429kl*eRr{c>x-pUa*LRPG2*}N45d~W3v znD~=Dr*Fhc)zq1Vt^F z#O2~&uv*4pJ)_mppI(SDZBcE-f=-w{icyz`%>||(AAu(P@mE96ZVn{R&i_1ogqJX&Ha0-rq-l*?t_zsd;?s}u)u z4-GW=+|j-K*%Jae+%5;k=&KgRbgXr(qc0S;GYOL$pQ6d-_IvLU58^zUoZCYS(0+)I zBSkcF`06NwAeNE2*_8Hy9QAi>S8}qnN4D@`>hosFG{cOAp1I~SpS>FRi?RRMJftHK ztOOw!tw9B`)%%C4xY=4}XQ>xEaND1ask}35Giobl_}NxwXKZ7dZYgg?$AXHXYm?D! zkA2K2PJiJk!a;AY1y-z2v&R=ftJ9m#%Go9veHyrt5l9$=%R5C*V7TF7zC~2aE)Nc+ za*QxD0u=YVaLRX-m_GP!=&lwO8#0#@&5&X5!RQO|{4EhhVUq(U^Uw&-m5H_u`dY z@c_fT2`Np=#sB1(lO2>9V`w2?3M|eZmWzT#B#IkOb`{@4^m(ax+z79y&8=WceSAj= zDGM+_phHmQ)>6{2)eF{)s8Nd9nuoas!NtdR&00@Z%&Izh@{Hh}Q5eHra36xqX(40q z?TUKJhub;MzU;G-bYd72#=@pORB9X@5|Gge07`?KBH{l?d-zwlp(rSM3@J*H1r6ZD*FluF*CaWc=diEDne-gFqEMv!f!!Eyf?@kii{;*SIs_d}* zTgPdo2&4b+Zb5+V@k%b6d>)^DJeV9tlpP54GM5QjV>@0nVw8l`@(XiDY?Yka9}vh_ z5qgLBLvGuzGwPck^~b&#_iSy=TYfaV>%$Cobe;_>!r?*w6CsX89%N({K{4qsow5Ry zYYEw$5KQ!`w4@TBY_tuC8kq?k2IJ?V==_w-)RS5WDq9429iTDskW_+d+Wo_2)Hrru zde-ct{5s2Fqc@CCEuCP>WlysQ1w=Br;g*Kt*hj9##=<#5zw}z59lZOcU>Gh~7Ua9-7T&Q0~=wIEU0?*ITlR%Z6m_kCu(b zpe-7YWJ|St1#MsC)H|efHU9TYMGE)zdqszi8^uL#+9Ufa@a1>2OR~v2jYu^mNLe`~ zNx=p#_L4CbZ{ZBoT#@*tis+(0Ja;j9J(r~%EL;&O?)DSSJZ!hFXZ~aVVl@;*`t~`D z6a#^{m&*-8EqxgW3kLp7F6@QSs3M_KDyAA&9rt4%A;c&5FU#%=bgiPM5{&ikqgH( zF=f(qG>S^d&VhrLlM}s2q z;NjMwkyc<%TiDo2u#CUZQIe^P=(U`s?lE|WkWu%9@v7|0@5PH`*~tx;A4kKRTK7I% zlsJ~(lQyqb_*bY#!6Be__4_a7l$+7K=$IzMJWI_bsX4p}7hgTS-U`vr63doVhARD# zb$4sc@<{=^{jX7xm(Ko77A~=PXZVZNSri*e2#y880GH3!l$D-N=X~zac;QS?J%8uU z88=H@Kx)9M?7BP*oY7g(4bxmu`+9oE!zZ+d4ofEf)vEnm5~FE`uW803%uF3PB|lYP zuC+VdWB$6Qd|S4}b-crLoNB9r5h69~VBSf5!?xEvv&{`+x#jqWVa6-`YfW51h~Q6R z9cTJpxzcuV70ap#<;rECMG;T&i$e>3C#BrwTh`7UMjcr)Z@Q?7hpOg>uxYt|`Gg+3 zxt04cELt(PO*tK$0 zp}j#}`dl5-)tJ#6uCt!3FAuE2DQ7Zn7i6sip4dk*lU-H_B&!Z@})~eYju$Y@RpA7 zCntk3gGHso@xH{UO51PdI~9Lj@$XPqGFnHa!gy$t9K{gc`F?w?My0mk;${R2$F`Fl zWQ4KVYjj05Gz2os$Wx<@P-(}{QoY#4*3x&VVaxnmg5em_8|=8nl&EA^!v2&X8Ta8y zlz=5KS&=0Vq+iC;!Hv{aF-vpk-ljcgD_ zDcoMsw`Hp$R$3IRfww?NF45!6NJ}h$oM4=?EaEf~9%Z(qaGF#G*%$rAK%+IN;e{}u z4(6S8j7Tssv_-6!HS+Q=dhzEnJW2R;4E}MH3V3}HSZ9Oj)rM#9(mqwVtZz^ndrcQ< z*nR)AgVm(dO}Y9R0`D?gIj~! zT|F1Co{Z6Q+~k3wdOT9;h)=%S)0w-(YhS~H8_Y64xmp+OuM=yblF)Gsv-@+kvy{*uJ(rCs3t{(|lI@@7Za*Km-Vr!JAV?Dge* zMRWmu?^kfpx01B8IJOh%8p)5yleha|N594L_gu?C;PYjd;kOgps4c(P9(gP}6Ye~! z&#&X#ck_wh>e>(T$1_RM%IWjkw0bMKO?EY`Z|BG7^F1(-A3x9L-xSj*<~aVI2BoWh z;@(pTb)`J{xL4_YfS#t$-n)Y8D88}xR5_^HGi@_`r1|Ar(tZz2-Xg(JL`K<356@C} zZo3p3A{cj%1CNXso}F*e5<>Qp7f$g|TVmo7@ln!LOYy6O%yyjNrlcHZZBsI7Z4{J6nXu*Ei+W!-+m7bR$Dx zmlx31jX*l2kJnY^&TB}=G1adXgZw(k2)E!UBkBMlIC+xDKIHy?r)!C5W#rMXzsfKP zcVwUm`XFzF{r~3mOQ2v84PCs4taX_Z)(`KGRUJ`2G+8x^Z&P?8Y9ezr7xi^~mB~2I zw%CbAiY#pcPZf}+Ki?N>Q7n-ye8gGDT^r&^RY`4WO-s)%o3h_&DR_3XI~a}#8M9!# zSMM|IvsAl|3O+w$op%GyVXOme%AeVm(8`Kx z7IA+)mB2&RFWuvafhLf8mX`j)tX-k_AGh9Q!Wr*WVDTlIN(CpbP&O_!)?$eQc3cN` zU=SPFmCPi$>ybUS@6UUT(5RwG*t2FM?JDf{c+dpsqhQKYbRn3LUt&Wki}FVb&X@`4 zNU)fFX-MV|Buc40-XYUDBu;`4LkzXWb1Peol}v7Jo`fh9j{eKw(4C`&|(00T&7iWOw zT5R=PXSXq+*igbXkOD+?3sT&2gHs~BJ$QxE@(W$+JYV5*MDdJW0ItKpEsM*HSa{2z zTfoq&p{qe2wU>2RX{(KUrdMEKV^s6H7F1viD*__eu4vCCbdeSu&PyKeLachLp8(Oi zvU}WCl*HZUl_$%xE~t!_SdIB?Ro2>S;RFTy4Wjzyu!i`VH%l#Fv`4rS(dY>r9?mWe z^BjS_LjYC0g|W!u63J?$)a~B7GE;ZqvuGGZN)BHN^AVDO*Lu?H1VtS(DWy(GOdu?r zo##;ctBG)}ycc$qh`G{g>JvChhb=3>~qoh21K_IRWcxYG#aXy}af`gAyMmn5(1pTn5p#$jsp!#yVN(0!R z0<>_t0RAODrSxZbG_7-%7PZnffd-?uLn0ugCpuW@iEo5GEl(PYKf##QA#udV9t?~9 zUlnWpV7ro_hqGa6F#UlPZ>&%5*j2TSO1X*G<4CmLJ6JxC9~r+^^;Dd%62%f3n?*4# zjd;?BCDus2ZOzGS*klhyQrJ8fiCj|UETph?O3hCDK$X!>co}nn-ZS#;!JW6bovZ-y zo-=B4b&Ex{^&$C7z=QtP24vte7IT(wxxBk`t8oXeN?R1Q<;LZF^^u!4o1FzDE7v=m zBoh|IGX1emoldJ|+7A3_K3O?e1qPGWm;CAXMsh0*t#Hc9l!iuq!L~vhP9W2iozF5< zuXmE-XJnMAU4NHgpA}c*@RvN8v$O|ceSr0R!-y~_p)e_> zNOVRkW3rbLvO~*F`dW8LZ@mXQFtaB^Txg(Me&UklFUty0e3C$LAdNpE2(D=cCV*p} zoBYY)fMv!#N+kxON9Ttg=z*#3H*&K})qN5t+e+&%ZT<_2u8l%5FK~`)#Hz$Py75dq zv+`u8CdS%~53K!{{m2?T7y=3)`BzA;pgVJ5h752Jnc;L2;J_&$*|V#0&jR7VZ+s~2 z!z{-C-ZsAaL9X3jO?vg8xKC!s#0;C;e^sp45ZdJq8EF;k2$v5&J6(||gBX%IqM&Jf zCZ4rx$jf_ze0_sXngB~bKS>*BIb6edm^U!_s4TRYwzc76M=G)FhfQ+`@36a#NH|0% zEzGh^Uv)htqm3zZ`8KL+2kix9!0OWV#h_@cx_vArhOF^sv2-3iBOS`4eXJ*&a5lfE@Inw70fzzOe)X)`8CDa&_Cv}$kYvcR)O zDQ+s0D8QaK%j?Yhr+jB8RxX`I?ud^t74TJnnv6Jh2krI`LkPV+*yx?l$aQVZ0-YBB zNPVNX3ii@W==Eov@tWm~+NPn@Id>)Z+omP@v`IzqMJ~qs(m1RTD9SWf{k5}xKnSl6 zYqir)?|ksuuAp-A*+Uz2e&st=mK6JUsd@x6zCaZFLXc2xDq*3X!x`y>VD*&m~o`!Xn9h z*G;{Xz9=vO4d;_`Pk1@hIoL#!D0O(AQ&$zWoLZucV{jfs_)p zB;t(>hqzDkkFTuS zL`@=6>&jn8xyMrpO*SRe!pUc=!wWzrXieo8XaYoX;);6Q8a`hJRAPCELt#HsybPh> zOM5t(>*DX0b2kGQ;{}wK{p(M?LJF&eyV2O^asmLo!_JP9n^xeXrOj2FAX8q+l|N~i zwj$}qhRuAsO=LQ&#`=!!@9w3<4fTw@+2KJpYYk7gn*w1GoFeZre{|V(MX7uaIsUVG z-3fiGgw9gDLBuMswcwJa84Gt?0=$`5N2_H0do>^SwIaCmN&55GsTD*^vu$0SdI7b+ zt|%BHJAAHH3ey=Dfv*>gV%BdYWaAwfPsKS!OAH6*O_jDx8(7||8oHKo<3HA*aOdJ% zTCxIgw}J0pA-LqWp!fjfsRHdMAM)9|q;?&)K4!Y~kEQVR_JBxvqm;84ZL>R0?v9J- zKo|Q_mNk`=k!&=Jf#dm6^UP< zuX;%#HFuL6jF7dj2)}Df2Zt8{jxQj)4wJg+xuGU{>ednJUt3W$YN2buj8-Y-;?y91g+rEEb(Tvy(r>bNwg zp?Xm!Cjsv{IaI#us3e{i<<-!H{K+L~;YK&(HJJ0_y%?-i53d2~nJLu?w3C zh;;u6#YM$l$hfnCUq`arSOF0^XG>%rE7j4%e!CzB5Ia&9C_@ z1ZO7{J;nWPwC*AD+s(lYdaVb;i9G_`lgl*v{AVP2y^%&gJEXwKANo}PJd%O{b0tv4 z-@YbA;I5*vw=e}Yu#Ifi%8(s7FMw}xOu+mME@Qq}zo_@DwTJ13l^LBb9DZg`kc{-? z@t)Xr`dT@9H;bo^&sO}O-m%LAQSLu7_b7{n1*_`#ve3eSE+QG#$sxVczT-0)yQ$0G z?P*}Q05iJe(t`C{n3ZD$kjky(pA1GHtAN*I(cPk|G@rX#Og1vA&YaGUNXdePSv<-v z+n2@vlH2F2f-Gssic3RoH+eoY75uqXDj#xD>_9stn~PG!RXM-%og_gD7zUY&^X_~U zSVXR^6>_I4yaR&HOZmxHQ>6v?Ut<@7Xvpm<6~mYn7w^t9g@_Q1OA6!%4t92&>v{8| z^A|HO$Y`wO&5+i$gAcGIFB8)oMiW2rbO?AAS5t;PzO0J$0y25md=X0)>3rQ5Z1H2U*JeSp+W<9dT?+HJkQm;3aiuJLuJpVs$Cjls-L=k&6iA48n`!EU#TyzFAxeBsuZ_xwv)TP3f(Z7CiM@^ z`=eH7m?>;K068oE&Pi|IqT4Jp@425zp${jx%B6DG#RHbtqPfIxu4RjtN5>yyW7bfp zg!!t}ojJdEeiIFV>Hh_PIsYg4%gn~k^1tG5k0xgvO$YMqH_Zp4be7sZP%t#yKZtC) zFXo|uVr-j6hJWV2y&h;$B8?0-`rT_-5i%GWuWBN6eGHDs_j%gcn%VGGuLA=QU)W>5 zF?0O)?=J6~amRJP@lj}2NP_NtU+yCg$Ew%avAexLPN+g%a`5*I{{OLWA2T0-52vrp zL2pzB$IsW767>M)`0kp8?~kwRJ$$$pHn2uth1gdrWq7Ob7lZOwb*@XQY_m!US|0aq zo@s~NrQPx0_ZsoC@2}f9Kf2NWixf1#gSvXspT4)xxZ_uom4TrLu>-w2MN6C1yeQxf zYDm@Af!$O;hQ4?(Zix$#Och+3AiBh_cu#P9MVU|h8YM??Y~}Pcv09Vl%vCN4x(Q&u zk}mWP|8hsUMJ7c0l&@aP1KndiQe)j>^kh@rIIbxcYiud5bIO9%$1hU5wPyL;(w=H2 z5u7%Wt_b3+E)u*B2q23E9{y;(8qiybatZ2tCbd34JB7H*;o1Up6LJa~N>H*?PeAK-%Lo%6K;!Ou(&yHQU{w z+z+jn1i7YyX$YbEAWn571U$JzK{NDt$n!DNLW)4ie5uU$kq~EAjP}KwS_z{wgV&~7 z4*340_ajek9^cx#D#4r_kJ`uk)hlru%_HhJE8)`_(X(B&?+z7a(Iz3MrAUB?+I(x2 zzjgbtz%+sMYCQy|4z(qhzk8Gr(*jL@xgfdAlQt=mY7L!+0g6NH>(cuN9-Qx02{wYbOL1ould?a!ysKdQ`JbpIEV?S#LVHx83!Qqgq2c!i2iFw+0pE zorzW*vvm6WvEiy%%w)FX_>n)^&z%Io4l&%SAy(gg7v~!y6t6yKC`ul z`9JDGX-ubwIUB1!!_|R0;=-bdYBHnAR_V`eeGVTO#H_FG-QupFAIBb2E2Qzx0@+~b=l51 zxqhz6oxtA*!-4&XGB0G}jBU>?oz@mIv>ub}mo1Jz#JBCX;v6fc3|M80OiWNTU%3vR z5Q{v(={RnLv?qL$&7{W{bU^OBt1IPMa(kWP;4;bYcKQkp=|i=ijU7eQ4;CGj)c9C% zjJFTj#Z_bFyxge(Ca=Flkl!5jHd3WBR zC8wR|hEoZ!Gzh^!)StI;H$>hs%p#%6rKf_#7JrrV;1gfl%b&BAN0@_3dW6&`6-O4%lE%h+;*aPlLe&-5hSzw32{kySH7TPpjss`1hy0aT`GS^A61 ztA7rfW?wlsP%KZ7Gp&#u6LX}6gKqHNkJqcg2G>jemdhedso*23_W*UJbc7=HH`*Z7 z_IK;%rl^IQO2Mjh>Eh;R_)WcsSNN`r?Cf7 zDYFc>p$nycdps4z%bML9{-tQJPRGYg+#}nSaP`_uZKQ6at$c>5*}+`$dXAjG3ZLgt zvWEVzlkU?9*VEYQRzuL};O$let=&;^^iVojMG?ao|F;X-xtOgwI_H%H9gdp4bu-KB zI>^jRH3PNuik69VEAGN}biMQ>nm5HbjbbetMnHR~%mFw3-q;u(zcP zw|KZys>VbrIctI7E$eLPWS)coheH^AE?F?^-nW|<0{nz79v~QT`j7E21jBw$M0c$v z0)+eA2h2u;yteA9OR~EZ>x}Vh-?!CnaW!4v|32`JxpJBu(Tef>>VaIypdtnS@EX3` z!;59n;qU2PY=mBo`G6wNhjFfr*DV$U%MYO%9XGr2J=}U;uLgWw)xvv)y<=7y^$$q$ zkVaaqt~rG@+4H5=ycyHpUUzbc&Sf5RWcBK4PjEIuzu^x>4hKe4xM+t+>32S-lS}Nr zsK2o&nl<6;d`}&D8vO=O7hrnUKph%`Lw+CG7O?($Gh(cr2T~wYnjbk)(h%IRlpZ5$ zhHhx9abnq`^R~iy0&A7ke4~OhWI@}rtCjVNVtrF}owIZ76#D(lHS;?-y52M|ez|1! zZSG7Xhhk)P@~*Ix5#xb#9^bn_Q_*?kYKgg9v591V(@iL2#fg_5pWfNx zy*To$7Ix(B@xW#%>au~TssyQit=s7Q?@b7j&{a*^W;E8fMtaOvb4K3=W9paN{s;`= zNWFM@{+Ue-NOObbSW?IZ#z5fxI1H2BDq1*A>n(=61Q0 zwXI;w?XN?-qlXws*ols*Ht;8-rs~IiG<$dJqkS-s#@$>h>S$MS#RqI?5nv4Fk^+^4 z0nzm1pBh=;UaOWT1NvvnQj=6@`V(mwm~+@5Y_cY$nT7LN12u^=J3Vq_2h-~DcAW3( z3MOc?VG(PBp)}>@$v_GEoUu8sJcVJduvx#oURj(i1`g*xEUMWjx zM|~V&o2PJ&IN%nPtsONk%a3KLGj-IML0WwM^PD_MbWT71)r+yb8~aMfqpfdk1#e4M zwMhlrqaiy>Oi_Odd0&=WX)&KGfMZWo@;UewGQVehApvhYFXiyr6tgu+tU(eBoM<0E zL2_@9YIpE1h{QE)&OfQbEE9`Cc@|E5kiM*bipu1$fsn*N`4hrShe~U7<-Nw5nXO@VFTM<_=@E5Q@8^LZ;&bia z|7j`NnEywXk^{i>zgbEx&N$k(|IbodN4)~W0siak+mpYUga)mz(d|l{UeEKA1`TriBteog<35W04#_j4~sl9z^^Y_{MeYbw!C2bY| zD+(tLzTQL~P1$XE#qss|IV2B$x!oAlvkGB^_>p|$csu*xjtOJ7U3@$ImdSu(Nv@M4 zyoxvz){gwT?)cNoG*30I9QA$di?8|_kjo+@X1&VRv za#pfevuY>xL>s9FF0n0S^3-tFkj-2vO4LWfK?V;BS$em5DOefPfbz@V= zKA~@;)>e_yY;n%chw9yo7p?;Xq!Zh8lErh}D|=NUSurvZ@h@B6F6Uei<^9v9_=Mh9_Ua0&P(vJ( zWtGHh$UL?2vd8+AE# z7}ymu9S%#mt>TEeM#NB&o_&jQyI znMREamYSH>_(zb(jsmvRtaA!_CZFQQ&QT-E_F(nCssdGB{c8;&A-sE$i+l>jF$^BL zy0oWc9#{->rB=X@cdAg>AzV`BP*L=Usy}y6MqV%o+-O@_BcBCH>5M9SV{1VZ^K|D= z)b%k^EK>ER(t+$WW@j*nAmtZY%wC5?I}LVAXo z+Hra==jaevjY+psMb_6Q&q4M>_oz;jMiV#uQdQvQOX&~%v>Kp#&YG+6K%ilVIGLx4 z(3XZ2c-alA!MONc>ZWKPwaP#Nc$=yuy4S;{5{q?Snh+KCMzhEJWi!)WS#2V>59+r+ zEK>)UqDUAM4=lS_f9U9Oj-0wftG=;4ov`wx#TU8L_zzoS7We3qKsEJT?(W||m!QpO ze=S>vq86ItYiI;vA0f||2X%{^de_gZNe+sx(m^@gsR`b|VmiJWqh+OJDYy~X>PhVk z#v=^$1v)n-UppHMk~1T;L}O}CRnWN}77RZzcx?KR;PC}DTh*98i%|LF%d>IjaD8t845Z%O%a{W2$K1BL*a&c8;`6$gc zQ%-!6oYm>6nsCc~9CXGMMtQP&!dDJgZgpwprqMcGibFKPkFZP;ey3y763LhVemLw` z;lG-rAvz9cOIj1{eq%KOL!(XQB+j+^t&Scd;rF%YG(yKXbDL`>_OG3XueN{$SP z8!aNjo0sZWZLDGx)FUPo`?k4R*<$SF(NXtF$_^AVZUpq=okbg9@r~coR_DlN%mtoF z`!rH`GorCJ3Sr-zwaq6z{!7vF`1LO7K}g6t$9@UIc;riQg>ZcvIO|ye7?`sp%nE%Y zbyPIUsD}nH8h<6cEa%5I&tn>g~`aSoBbpqs<@l0!FyDrCC^c;TNhDM%{i;`SBu z%bIZ$PBwW2ZSe^-5j;a51eVgY<*&YwVWvb-ATiTx^TUIHlH`e!mB#8Ku*M+oUJ%v?GQ@0;`W zZNluDn1JIDZ&o<3*@2DUSb#Uuwg&;jq1%!8L?=CYNQ)U1iI-CKc?Xr*$Hoz@i6nqB z=!l~_YFV01^iN+_6HuQX??x4p`cJ@?z1t?KC2 zhufuAc#&)(Zly7@j@={lhMY-h=8_!)io@O!%>?G=p0{uX{^lJfvX|5XPlqa+jIZ3| z93=G2;vD5{&Wn`((aQr+K*#khr;Pv#>|m}A#*d!}zVagSAD;y7QDe=oG*slgX#cjm1;Alr*bAt z2NgL7o!AOr_vvPx|9TaOlwe-!o2&m*Tv3`MrtXSx6k|d@_}819mPP0g&NG_3lG&Ru zGBr&oxh<%GDvgO;As2e^MZGuea~X}P9I+JjG<WT`kT;8Fl&v+Gq@>7Y}vFEJOhjMm~04+<@ET56can(9+@G z;@-Z88Ks~A%W^~p7>H$155WAW7ld`m&I;2Uv?^SR!zslPj@gg{rjjfb{1I>ah$vFd zA;=72N9tk2XOV7K`%7sPBbCQ9^VWxrI(*aS7b8JWO;nd7{mxZ!u_yz`zvywcj9>y# zL$)q~eTqrqI9POlYL_2KX7aEmiAEg0)enkSaO|3)f_X&OPu}n@gOUvq5l@f`X!Ykw zrhci=qQn0ZdTJQWHH3FY7`|I>)A-EZN7xh6t<6zc3AbUb~XUA?GHE@=b z2x%HEG)s?1MKgxU9a+wLt0DEz|3$kuo}514sbUSmIh>)J`&(fFiKI5CFQ2&b^}q^V zj}M_5U4NP4=Bp?)8xV4JIy*!f+1F}-ir!6BCTQ3Q-HwR*fh1j%H$wjOOJiANw-Hot zXcU_D+KDg-1P3GpZ0KyD!w&=9M-}+6n<^&5VTx08xfOXe8h00^)&Y9^i@}(u*Sg9_APc+4e z)48uK6~#Kht<$dP?l}%iSXipQBOnVYqF!y8JxJbnX^UC%neVs`3I%!EO^d{eV^%P0 zw_2*-DNl7JcRWBSvO{Q1E&YK6@*yg>104CI>f+Kh4knuVB=9bSZJ$?V;T`&-u+&rt zMo~>+>`m7l`%$l(Zw-87Rdc|{v#~1mC}LNbO-zC|Dr=I95pjab?xm-OvdAlJDkFg< zh7?3+fJ&f25!|UeHqssuqo+^cfRA=NMqN|>NJ5$<-CRWGO2_LPj82(+N}dy_j%fbZ zt5Jd=Z?}g_IUlxXKTvP%#bDH3e&r8r+E%pcd~HUPvC|Crs@*Xxw{NtzW>r~qj9D(KPDY(@L({mN^uQSC!Xdiu3`POeWV#`9(w#EI@6dW@E{*gYCA3N38l? z53_xB%_Npj!)i{EJ1WazPD)C|*>~P0=w-+2q;B_OMPe`P`fkLP@jdID)9#1opv{NL z-KImY9hD5XxtL7mse7Xo)p~e2Z!C`XSr?6uXF!X#*9hRI-?F)GE@K;9l?>V3j?o4~ zQG4!1xU#+tm$q*b6-CUF)c%NEd8@9m!x8gBqB-9zlxCOI7|zpC#^BEWs+sRHM)dy< z0RadZm~HMCscHW?w%m`%TxYuu`QE*((tr9x`0 z{q6Z?!31tpQ(PlALX^>ch}_*Rxm~OZ3_TZ1(3}02;g;AX0&-COPlo$j(pySb)GwRp znP9pyGNB(ys#lZ)n5A6#ee*BF1;bQUr;AOPEMuu=xiBh!3HzW=wFFmVvE`aJ(dv@? z*P^8{kIX5Y2{<^6(LxYons#c@LNWx+N*!rcbyD`B4(Vf{MFeuO#1 zIzZrZm@g*+T~*JpJNriLMxhm_(%0By zoJ3bqY2kh+@QVU*l6F3l;dC2$;m2H;t4^_luy@}YX)WxX4u%cNEPCj(I> zIr~mEcSx006AYok(BgvN6$Y|t9T(N8A59#)l3gB+Ifys9w;>EFzo7Db{YUEfA|&fr zWlG#5#3)2l-aR9_?+!I=vXX`DNSQQ{1{a*CH$+Oxf6URR$qi!B_(MjB*o|MAOW(1^ zQ*3c+9x3=UMCIA3Jf0`sKE_C>%E(vMK0P_eH76I@%zG6yQ_ZA<9c&)Sn_A zJYR$?Zm`+J&22W(^`-10(N$S*{e5bN_1{Fp6nQYCQj@^hnu63lK-uo!+-M9Q>qKy# zWL47iryM{Uq2QXJ2?1=3LX)Bc1}3-udB?g36^(g+{OU4p#V`w^Bc1~;hr;#Av&d=p zlc+}Hf!djZeNMc$T9@u)O1dUU@g7iNr^_QiG>rV=zdNT|FqESe8!K`wG+Cc0B{QAP z>IN63T`NXNCG_BxRplK14fC!$8}3}OT!*No z7NzFr;dWOU$tK-p9+t&ef1Ha`-4&k!W;&FH&zWjfei;3{YK2F`d490In_7`wsuq^r zjL)82udks}eW#SQI6DUmjT+t?Rw735Cu*FJckN1@A*+33&I>mtj^MyRGdxTgmJ$Q& zj4}9(7;my28PFEjWl!gw#e zH(1J)5h>Zz-C%rtn-+#O8wj_FV)8D8EyI~EjYM!-zKbwG0`ur6cHERB^PuL@MXuafAf z9~T>=0V)9np&ydG*LevK)oqeKB+c17H>0GQiR6UC6T8qTQejBS@Bf^Si78NjFo7`7 z>imSibnNt4;a#6Vm@)CUYg~tUB&7+EsQHA<<7E2e#-|=7MhYG$;0r?C5mipQ%9k66 z6L%ta7LN1JEd$2BZAEn^)!HTN(=;BSCh!>l^f|)lQle*|kL*Vm=Ril=c6+-S7${W& zS7GW4m~#L@BO5iw^he=l$WKFJT;s^qz+C<|;q)qP_5iK?%v@+tLT#^TR!pPA{2etM zfdjw;_L{D=>)Yo}ac@c>huPz?n=W|R(igC2U)Z9bw^6vyAqA(Mb!qV={aE?gq9r&K zz~@qsg#SyDg$G!q)EM+e2b91P@7Q!!5ef|Bbf}ZD$|OGkb+VRru$;jTwJtC(<#rTy zqO0u3&D>m*jdI}gE?mUTR@Ybod76m^9m}~Uq>h1D4b3o$`~Uwvd}CeSpMu}IJ$UFj zfp(#hu1tLAYZ}+G>ItJ={D8pLYZR~(b-en`5n^@J!aM412qvm)C)R$)1=g)}BWfn; zstBQ?77cTAE@pLCY*mvJ(L}OWzK6~F?B!3`dHca(uF+a%PvWV z7xM*c_KY;w`(QVLuuUeeYg`w~GP80O+i1ktwV8nuECDWi9$pZsMOHkl49kFMQq-os zrkj}jh0mYO63mIrYS|qxOE07Vo>~qyFa3C>5AW`I``FS9@it%-o*(1qc|JF18J5c~1jXIyp z{Js~{aKrsa-vag`t4#~T83^(HLQo2vv%~92U{ycBWWj7&nNjo4u_C9#`6ElT6h6IzAY?8S%xW<4E?o%$Wp1)>W4%i0_KUn>tJ)h1t6 zytTdOM<0cQPX~NOo@DApJt*}YYqwSZIym5x+sdU3@=%{ZYM=&Q#xYwBlu2lzqvDx1^VrF)< z{gpi8)*ux2l}kLUSI9XV=HA$&qGLJjmzqDFq#qq-%XwbhB!z^Z*jd|~X=H;=Yb#G= zH32!0y{XrCpU8Dy^nY(0g4CRapI^3I)?~|GMtu$;uk5z^z2A@hECFDH{|^db|F42S zR)+tmQcr75#!#~$Z9k$u182V>`bt#mMuCVo3OTtIYABy}4}+mqy?x(rj>@W)KnDf2 zm0vg(``{qXrR9hf$4R=J37zwCak(fzzr@bzx}JRiS&$9|2hpQIVB z?DTMdwb5rXi+nf+fDV` z4O=n=HQTslu!+9C55@1<^6j-*&~W)AzS6VJl9FiMwJmxhjauk}{qg*qy*`D8>@5-eZ5(# zP%`_@BTgtz#eFE}s%Ol<2lu6P@*_M^#T)z^tO=9B&>M1+?!41{@{#5!;i&;!+%6W6 zAO>lQLOtw<>05Feit$p>m`{WE^{NrujPn;LE12vo?=9sCZYHddw0GWvf$th|rwP;z z=oC)7gIUqVhLo3P+UXq@+PqRxq2+qvGA=VoaHd-$gJQZIc4_fCMub& z7Fi;LOGi5EQ6^sg#y%}q2E`Z2Z9wK^m&ve*dR`Q*L~xU}@Hfev2NP&wsqAo%YX=Kj z(OMTPW@a|(eS_SP3I| zX%>681Z^0;IwG)Za&}NA4f{?ejoGnVl;h<(6ZHb1tixM^ifyysAOvT`wR;ubz>MLn zzm?|E2KasEb3kd16;98QCPp_0RT_t+dMQ0$TNN5u)|EciU3UTKx&OFO-TEzfDBUY@ zvs`?etJoiU`xctsbUVk6WQ6P#)CZ5GHMOF#3u`A?>kUX6u*}NAR93R|jW$BHXL%B9 z%?>n;Ih_qx=`-0W1;*C*?zpq*c`fWzT~p{2cE@p{45ES4au>q=nX?-5Ko?x`$J~Rn8f?lHSIqy82z_Pmc8+(5QO06|50`lU=khpv8Se5;ZTWVUJ zbdI;feYJvoMlqo@kEn>5JF@XyW$o8d0pvxXpSJC^9pEtqDS96pLP-?QC(w&<*`4^!o^$@_)E~W~9kVi`MocQN(v_B80LlbHtR-84 zR)}0tw}@s``EGv9{Y7%Q`WB!HbaG;EtYeB5G=3I@7?Juo%7xLT&SrPo{rlqSEA|jc|N9B; zB{~FnjLD5|&yK*`zrQzVOj{&h&UfX5i0fT&anF>MFv@t|ZL<4>=?EOD+j4Xsn{_*+ z_e*4&{b<6Whk47ORon@nc`2k##7Gx_1|e796|fQi(pjMsf%(33ZUD4GWbZ4Fe@ftd11RQDsvJ9&eKI z`iQ2ZSqZ^fJB4MQR9{E&@QJ9}!<&&fp>}INiTJ~l|mrL~mDe){wrH{ zoS7M+P`_$ouk}d*c90Mo7plvjk-*1Fy!=c)E~^Z$pa7OCjPtq~RJ=@?V7QGkmeiX$ zybFQkksS`PIGrg6H0^2~7$gqAf*B)xyNec0q}3ZkafEkZ;kH+6exx#em4QtnTu`UB z>-&ot)^A&o*Sv_n|5yiU(RlXm3IZNfZs&os%2Xp;H4LF>WP zZxoRYbUI8LHn`>78%N3q81nOvhL(g8HO*ngW3vJ4?#{b5b4^qQJ}V_$C%Yr`P&&+z zEw4THW#PLIeJ&8ywm@#2_Q8Fr8v-7_W7`X@)V81`-1eb;DOxR3KFV(*%OjY97O_(Y zRn8qbP?h@KUOH61o58O0`+PUkO!(4YpFRHn*Cb>DF5RapiAbj*b8y;BYJ#NeBfct zfI5FBLnfmjj+Sw$K5&N;OwC`LY{md6ayi1bfM$P2!RyBckO&-U#5RCte}6OfTOg=f zCLzlvg9~2_10XCT&h>=NGe(IXxi0Qv=dLeZp)=3F0z+af z6n^{+e}^;M+e6fe`M!QZ&Aa$C%XQj4JX0!4FC(5LeL{|;)ac_!?$zmfCwRepb-l_# zZr7n1p2XYbRR^BqcGEOb0yXScDS8!@0Ci83wi>j{YN&L;R!$i8Do`Z-YeU}8(N_~f zh;}qjSWQ*G#XIu^yK66SOpW|^(kR$zNg4;2Pvt41U5;nYUd$|+b;{hvit4r&QQIQ; zc^QPPkBP~%*BOS=ce3zk=LeRz`})mtYgyL5+F3%ef8<%Bw}Dx;>27L;Cwt1`{(_~P zj6e?+?U%5%u)kJjR`B_~`YFEB{vR*TtPKAZfO4??M+`QMJsxw_a-UN>2OmXrLo$~p z^p}L+R&bcbcJO zujIM)>t^-qlj_Ge`{&8%<4|v@D20Q@BHZEQe(KgGO$LIL&*%HFg__hGEA^MNm3rl* zm-5&92h*GOmbk0+`1|8M-)4f#xv$phr}%4^(9dBKc86U=?jt@IDyZ0C?3b4Fw^i37 znWBzGV4}BkXtr{oG?p!|%vT4hIrsPLzLkP+X2h%X5p&<1l1{$F_vM?oQVrPX=m7@MQwDQ%w_>i z_#|~=i)kbYSu9TM-zjFyOFd>1+qGRqAgh%k8LyUX9Djw}cCDHOE!j3cS0s0+kn$2! zM>J6)SP{BKz9z<%I`2eEq`{r~LhwlBg}kSBd<)&<833+`>OUaasTK^i5b!}FN4>IE z0sS>pDO5`b>KCk=NAmST!fD;dwCQ>ynJ6ZHoOx?s5fefF^a6=Pi&4zGic9s={WwEX z6=&=yc6IzMw*;+sRG^lGmUW-xQ4{bt2*ak7D5(R^K^W(U`P95g!qzAQpZ*yZQ;T|u1wg8__QxQ_A9>FT6 zOBA&Mc?xy`zwy^WFo+826Gx-3Q=2%a8LLZ&AIhWp4;l#}oC{cbI%=nT(rP)~lZLIN z)5C_A^TE2AhK;o$d=_Z2l=2g1<@nk%RKP&%d$b3T?=PtHIhjwh-rJvfY#S&PUr&N# zJ%I6x2JAbOYLtYma%#8n>8JYhn3it!YTkfRNS4|8V$8+#!Bm{kO#I! z?SgAVWhWJLP`hEjnNMoXCCVKSi(gde^&g!!rBK;kjh!?k8@bZ1^h8L8ac@bT^{hs6 zyDkSUjFIL*Xg ?Hb{97Cgt5uRs0nPfERfZqh>%sXmkaUda z;Hq@5Y-?4?8OitIFlz6w>)ZqShv<;n(8J!`#A``I3~oZ^lZJ0?XbaOB?6)Pr2^^zd zR=Oo35i$ekugK06j*i5s={8q5QLQm)xIO(>fFT@DVN)Fs&-2{WpZm>~F>KFw=_jt^ z#t0mloM`GUf7Dx-&qS6F7N{Fq85-bAR-~kF8Fz=*F;amT_O*g1Voa7 zi>*BDK0@-qFg9AQdXcz}g!>OTdY=#ZOf^e|I+F zIv1Bx{bW!IrGQ_TJHWVT@BQ`{rVbp#T%beGKR*Ys49q_yQbuo*}3dZHgw<&cj0=}E{YtHIQ`n$G`;f4S;gv& zZ}QPp;3S2=;O6X7TBBLWO4!u3Y^RcWK|1ByyC_XBViulZ!ibQ72S~hJQn%2`t3Kh~ zdtQdeSumF!!F(T9;5W?LhikTWompw|mPU$cpnCD>`*EPlZw=fe6E49%HsU%nNvqsj z(+Ky`&1I4rA$!8q({*noL>HhGKxY&VW|vH1J&a+B2^HemH((WU2m>P*r058Of`~A| zxNtT8sAP1vK2M|MS*~m@LMbvYSQKIwC>_rGRCwIO{I6Qty{PO}WSz-KuiNQ0Qjeva z9k>;tp;wX8NLt)yomX0u7pcXi-6msOPp9S5aAU*r1Zd4x7itYVb zD{T=bBlT;0tIR5nqde}=a1^A{x(M+~-6j#9c$NY|4esh;XCN2j-M4jEA2_bJSW+U9 zYDN#T9dR~4n$n~Ofga2V!nh8CD%8-)kNEG^)^q?#)lnWWoA0@5~t)TQcS5Z%C4Fd-YL=xv|_v0p4` z&-k@pfTbS#lvM{e>oddY;JK&5RlFV^Vfbu&pzJk=&svNOEYNXN5C{a^0lZ4_X^5UB zy7Mv6cQs1Xn2Mmce z8m78p)MufI9Oc~p0eu}x3FGhwiiBv==JboM;LiYhH}-lhA=y<|N>0!f-VBmnyD}Yf z4cUjpqnBxB7qtT!^AG9nLuO}N93k#8T0+*m6&Jz zHmab^kTv?or0n=w|1HOhDzxUq4GA4bc)(=LVf=c|NhD&$EYmx03@#RbnFdRh+5jB5 zcL{zsWJ{6328c4wcKttW)N)fe;cfDpZuR`;ggrQ|SfEy1c4-{W0k_lVvQL`a3(_t< zwyni7MBfWo5_h}k$->8iASt?_ENsWAK8k5HwLh)PrsgvbBBBxuBFk=mG_tq9>d7|O zuCf{U>vfSV#`~BIasG9K0BJK1ImzdcT*~v5=Dw=*&K~TFrPE|*F*4-n9v8hlOlit=s zk=$hxGAE18o_2DX|EAo%0|2@JNCdj%QZ6SpgXD@IY7c3%hZyaud4H$SX#o#k5ISv% z<2VjhYLy-i!>h6yOg$H2ZzR=q!_{55 z)-2@dy$v~nxMC13WB(?}I_FEjg>^!R-=ZsiU17G?n<=RTJnR#>TVsWy!T#ycmLL&> zjOcqMSm}NZW~)&S?!TXj-ns=C=;bhptADLyTHR5Tk`04-f15!nBwRsGGT0L>70}ED zwwd=-t>R);$a+)Pj;8e2GMMcZdA*pov7dl5&5#5@t+RTJxjeP9o67*eJK3e_+0eE3 z!MBHZM+3+YJ2F=k{+xZ6UHssVaG?(5%w34WA@q(Qq1gomfXelKUI4WfL=u}s&N;+V z>@KdR2%vWG=GVHOxFweeoK_;Mp&=D%R60`?q0r$wnG48hK-JB$do4$I2`y|M9-Wzd zTJeTeI+gY-mjOa)61hR$q|*2steCMosN1~QxO`itbESd9)@x}>DR@6Gi3{JXR>SzX zgBE5OjuGtd=7+S<;4GV%vGNwC&i*Fybtr0zrS2cE-L0uRVyeO_A>||m^Ys~d&yO9Y zVD>0%>=D(X`^@R4Yim=b!S3Cyrv#_qMXzKVLB^d*>qRKqW6X-K`NV znejgezgTNJ=BN$!-)?{l(2@c_4gP$1zyyGH8pc+FP|MjB-G)$}<7Xo&lxKy6>kCL% zY#y6`y36pRQT*MXw)4!5pT$;lTZW>~hN3gO%iGPy-ImNT0oZ_FlkE#7hX>NDMA`**XFyiM@6pGk{fW^+sfmyHsI4WFELuGnyF9XFKNd-X>9 z_uK2M-&PCXEraD{D^E(b#jAnQdpG>s`D39ybo< z^PAa*FQe?v+yvrs6^i6{4Ogn^-=lkHRgWFXE}qM*@1%!tZqRSJg5Jr`=G(|!OHbr< zs}n-r4X7T;fmE*H*YKQ9a#vjY=PpZgq^ViKWMhPcBNvovqdnPAZ@X8iy4;LTaI5A* zqe|@Zf_1t!7U=k^<$&Ri1td@W?3p2zor1OpWHdDo{7mx)Px2hbACc63wX9Ke{zVoI zi_VH%zRl_M1my1tF>{-E8be=_ertw$I8C! zriS-d!2!z0%VIC69f{E|paNo;4qSU>8{y%RSfz{t3<63KsjR^n$r)vTzNF_{vpQPlZrt91QK= zrc3RU^GA;sPND<|T&?W=92wtWwH%GyB3Jv(Wcbg{5r_EUcuXHw2Q)dMDw8)AJH!u| z(Yji6B&Jb2rqB9wFWxF3OyV|oXkS;vJVl8_C6@eqd(doi<>w4xI()Xj2^aeS_6Pwe zVRJjI=NjnFP^gHuXPQ>2)IxLrOs!1o)UR=@+gy1Dm6n@2tGY5c|9rwOWuFtk&fhC& zjVC=a(lLepArg=KMNa@}`ul=MY~Gme)PZrrp=#vHw_ky`yRFd}eHoYCD3EPcJYe~7 zXDQ1ak0(vM=g4|8%r+XSY53Z^_cioL690_vWcp99R(faGtnn-_!6-+PZCrq*swJ_M zNWsPyD%CMfoCFAGEH1A=@rV>SmI1_fE=jc?i2IBv0Prvc_r|cO2GEz709jp%`ku?~ ztZ=}4YTN*7fr2CHF)xn+d8f$dd&7cC(BZ!oV2&ubM2*yJMnfPFQgQ|2JQZD{Z_VD(sQ-nqCZkep#lD|?rP z>9HmtG>1~1OX@w8Y5O)MQ3Kn~YCRoVb)67)>YWi~{_I{)$GT<|@NE5gP* z6lkvAP7nbphuC=#7HXfZQ4sA_269=0;;F#+W$Yf=xs&f_Ort2%*yo?sSfcaHg6#g#?xWV|(4$X5vjvA4 zeW(5T0nHo1c~S?_U4LVU{>b%2?kBWLlw_;djj&BjSgHpOTf!Y#|>k?aDjt zJ}DmZGQWn!gE&iLW@dlgEOmutN}KI_a0}=v#N_KZJ+sU@`4j~kKftS)Xp--t&UD(E z+*Ak&FUZ~(R4L6l$=fods01w?5l#Ak*&;0qx|zX{H0sllM2zRnQ9~IYNeFi$`qB_e zOMT=5$T4#NoN0Z6ANwWcCSSLi^J*_Yh!SRDy6X(J6-Pp7wc~N&tXRBn?yd1Y5xKCl zOV<@~(!@pSbID9v)Wlp*CfBzmKlo;t2@`0qr1gI#(Gs|>Na5B%@mrd&MKBi6y@Urp zY=2xcHD4&;N{S6NNuBDchT}QiDVXXIE^i$(x!N}n9Eatq-f=QZdI8s2Hrlz=To^{yg;s?u`=1 z%^9qK7S=}D=R2ql=;cF0cAO|Ctp)yOC{5{S*nVQD^f<8Xjwt z2o`!WFZ8(4Xf>eMylL=SnO*=;C6HeJh^6EukVRH6B`@8lqZip)%G?!3f*31!P%fqK!_v|7aYB~v z=EyvbB&74t^!|0n7@V$>vKP)%Xn&`T#__iC;Sb=B2p zt(<_7cI(eOOY9i&AM0yE4gD^ffXT)k@6|8#l5K*$av)E~JePTY=>4QO_ws{{o^3_= z>lB1J3wy#RNYrjnKq#{jluH$Bsu&AtU+j;*Q|jRbAVti89Ncu6QsWl?UEcjb29Dax z4PN+O=hyF4POIw@uxJSo@y8I?xf&nI*K{LmZH|Wt;ZFj$iD;XMN#M%0(FKc&OC#4D zujk@yMZH##>LVS3%}g1@3zwCefJtQ>Y;e6SsR6SQHwZUJhE;9fhrXtK0$r%F`cR)@ z1YQ#IcKa{cf_SM-Biy9j7yi0wjx!qyr|G!!1-^LR{C{2P^RgPU;@kbsI$1jJ8E5>9 zVBQz4yvFu67mwCv-aEA>C~a!2%sVMg^M1-hzD}@uOW)N)TIh>aKVk-@6@&vUb1!oUn18!t*5{93lA5s2~e z-X)DtgeWx=&gl}e950Ddk2~XY{KLuJiCNwmn@`L9 z71W+^c5ZfIDtG$26RgtZ4cQCQi zBCG|JXc^a6sK+Xa>`dOGdof(|d5zyBD0QbOI(1q1HoNt<%sGa^ zD#=i!+Qq)?99Pb+x0b&8ZAaa|FOR<97BUbv*=_Qc`*8EVzJl(A{`h+NIOMlr(jL}o zeJ;JY+4c%M1+vXABzjrNWe3EgZv2ik_il?8dY4{@DSmD}e{fUwmN){(jMSDmeSX~9 z{r+3tRw+l(K}sBEx(R?i8q4R?`Q_XGp?PrJs{{UHS^c{Iu=z*W%Y}oKitg9x{q#8X z*8IMhZ&cBt904noVvFC6!0+>hkvr~}xhbo+%pyPh!@vi|)2MjpkUD0(OBZN6^uzKI zwu$Qny1nRGKV%fK4zBjT``ZXF_XdBG?^dU9nYviRxHSnxnad@2pQ)DK-GN+b)jEf~ zH($6+B0kvt%o&K=rSa?&qh{O-+n$)%h96B)s41!lhx($sGasXz&%?9fOF0P$&xcEZ zzB;WQ2^)GEC!%D%&EFcwbcb2HlCrS?(e2L{co|2^O!KCEj@}3|a6WeV?z%i0t!&FoApjgk*jOLIGMQ=<&}t$2X4|cU zi8a*v5-^+Oy4NopGil*;M;AG=1&;!0)wC}3p@Peua3&rv>CRXK8PzJv9*tCM8T=@P z&EKN0N#sT|Tw`~$kt%y7Q4JO5_S!;rM~!g4z@Gj1QEW;@I|da!}aFC?G|&CN2+Um&)7j$al(|PGvitUbzMsqbQr)IGw7}O#a>wY6bc- zY7xSU^fi7;;z^Rm4EY0j@pw+Whr9LZfO+9TDqi-adg+utnEQJQB{W}NP^-Q!V2e{3 z%$X9#*C{t7pm%CF^}@7BaMw|8acfVs)zrV|nsp2; zCC(`HhpRv*l9S7zm|;RbEFhXg>;%-D8cXJZ@+*O2*|jxnBr=rTAlZwimmj7ffXL0W zUhtU+C%hGZd3%Fdw4*7iv-s8qo`z|(&D&AEd5QqC;-OINvQV3)fOqPY-Hm$QnKb0< z;s;nBTevr8mZf6v_SK4DxhztrK7K7Zc~`7lSW}4?$Mu-9lUKHK9gh4ZK>16FIWWiY zSz%}y7G?f8(JjOMIIae}DOzMdtX{nDm}Ruim)vWcq}+u(Dj|iJa)v=4R>44?1kBK% zDUo*5>4aCQVJ!tzbVVJ>>}Sf5)xrikYh@tApH~b4dVa|YZfF$U4pX;-m9Wfwkes&1 zWiZ-G%qa1Z^rMxfB;M$jy{wZHD{02n(O!d!B6Jgu3BBj>6t=(elsHs8g_kA*rAR4*LDNYov^eIAOZ zH2|@vRepHmct7rFo#dR|&4OMF(j|BYd67`yFyk}p>a<0!*>qE}vp0jAi}I$3l`b~>RaFAc^tEM^5&Ph$B`w!Eh^;B>>ZG-+5(p_8KpOiC(Q&7lQi7 zW}0MU2^M7XEq_x|9u9kI{&Y9glPu9R=3E$89#j&wtCAC))ouFx#D!VRP5bU?|6`KF zOE7R|M*PVZ_=BT+@6ryFIp6zc+>w$Mo1M+U%%c~?zR{hQ*g$wTCsup1CAy$At{CVu z1l8~FkliOQw^1lKCidvNd%Gnq1zFToH<#i~<5(j-a!h@FOwDl>j{4Z2F&p5{>-^~U zjGB2-st&iL4`yXiE+cCJmyx~btptOj3vpQ%u_r3&upidfrvf;2c(5G0fm2&O{7Gb}o5wS6OJX zP9`D+L}&y&w>;EExYd03tecUG9`Z?H-hGsn-|(`y*#g0#G-VZKa`uwN8!tO#kEe~$ z*@PF-3|B;AYBG29J65S>1U3m=kV5W(43iyCkcj1|s^gnW9JgFTSbu#MqMMcbiAd0z zchAhX&2f_p3C3H2m`+eP3ECMBv?P>Il=i;2s<@+;UKv2=9xtC^JRRxv#dPUU3KP*f zaNJ=;ysLjY6MAlc4$vQ+T-*2LLU%aPFfJWcAd$)`Ctk$g83lZl{|jGz0w-%o z4pL1o;+$^I#@Y_qw_3v*u|SQ;!e%gBU<+?uG2Xse{YX`z+*4mAq64r^@*s;2PM(sf zgN$&*!A+|s@7&{VkI@-$Tk22oR!&klLWZivXt_Oa@LMPh45@-C_)_NPa@O<}=3cUb zAFQDP#qvR_H%*y>`~5T0S!<{30|66xtF-m5P8=H^x0&)eyje+6aZ0 zAosm26BE+LowwrdzA*ht&aWJB4Pixy{0swDc&IxJ=Zo6>JYHFNjL3cEsH|ApNm;6P zbaDO8fjJZVF%pgg7TMpI@2GPre$_kq@j*Kw03D&4cEc}+WL-f6mek{QyUyL~?!3^? z*Y4fXN$!*YonYhk?A`2_hfm-F_!q}M=MW%Ku+!kQF`loib=2F#)wr9DLBgd8bZny7inUs)O-%I`iN(5FeWw4KM8m4B zfkBKEp#XxMxi-Hx;7Tpqmjjc7z&-2>zPT=0ma`C59mIwbo|B z2DKVo*dSSMFIlL}BoNIeLZeM5q2`P1CmC{X{#TN?tnl)1kdS&eYS2%HX*=q~Yq4nL zcuy?5VD2B_zvGi$>_Gj$fY0{d2|k!P{^Jp5T5~++XauqQlj;pv=C}tSm=mK8=8V+b zHUyreR15twFVSx|z23yWykdnhfI-NaVRcMhmwNnsbzHXOyHUwiN!Eq7_=Mf2x2Yv} z^HI*zH|OWawBGDZ+=e)Z942oxo_(7r z_mnycMdYnITi*WuJPa=Bn&m1wUq`t!Z-S{`eT}A`uUfn`51sdBgZuv5olO7D7;J8q zYSd78&$lSI7)#gowz6wf-h?Wz6ZXJZ)(gTa6j%RE!)NlmO_sizJW#)fjz)=_2QzQfTXj9kSto|Y6a?Eu z1bH*0qxF4d$s!e1QWTx5N_+#aT z!<#Lehr2?AINS-EL(5!qnKJ5&9k)A|5?;bfx^UYgyK7WTw$Sx&es|^4+rw#ihIMH# zvn80r!wy*LONSx`+tI&w$&9ySzKs?+^wS{n_HliUjoarJm-i{`5l6}7Meu99NC*aT zlP7r7*6e_*o95Nhu49;oTq3AzmB-)xJ!|gJfwv~j>V?}@^8V;#eJUXb z1w+zFNL{=7GFC75T66}HFxY`(BZA5X@wJ?w+vH3sz=>KlgW&G_HXGtSD{fbh!8kt| z6MZk5Lu&{s38D!@)SDT3h^tB3dZFVwvztzMJw_3~gAcfj+DzGwdN2!^GADLtfCBVU zK$N{P%s5Tu41-QQh0O8T2%OEt=XkQfnx5Z8%yVsY*(P+zIaTLt2C~Arn-H-K_0TfR zuvR>QLW*mO!`33fGCg-!rQ2XO?@$CpUnM;3!TK~!c8tTJJ*$QtQmhBPs%4>uSuHp; zVWLR*+1neZv>=|%y!xwr#`;Jms(Zb4)-9k5=IOs&i2_GctR<*eB1SvvX|SIO9Duw= z0)in5?o8Q9eHjNl5`4a~7crY_Y~3L~rjIm>jy`F3qWco?m)4kXl3`%~9WIOpr=ftw z1AUQiPqapK1vr&Dl!iOY(7g`33@OPi+{ zB`+p%4c$AG(GOj)3vp0h$=3w_+~DLG0F0EHogAoKieWQiRTnd@j7CcM%*o~S>=n2m zB;IO-EoXa%xA5HIT7BFgKrY}$iOvz|(61lYO_R8>W9F}nRWs_pPk(v5;3Ze~;rjAC zRu1w21-xwbV@;veRB2zM&*|IAn>r_&oIo!AIv4BglhBNpkr_|Mnt69L9-?fr7$02f z3ri9fOf%&Zh$ai@ip%v0FqRv^9Dmniv0VvgoYZrDr2bV=u-Qf3m4ZAQ!Qw*I&Yr}Y zc>d{&WXz+|CxnA?5Y6m1!h&#{ksKf!aflm}%8;F7z!k#`HwKrxu7&RYDAQ~>kUbsC zrzDL{8az~5*apgR@|zmcFB!#Y$|p8rIS(2qK8UG^!XyIcy}X-?79uT%!j!MjUMb}? zxx_TVZ#G%sX>E%0N?1k-o4m@N?6GHeHvD zz4YkQaZCMb-TVd}ta)8GNYlN8kxH6S~5$oP!C{dw=cRW6w)Gfw#U zc;lMRLj%6~NEEeBEvh2Q(}4ZEB-1snqAv5%uRi`MPso13TRmsNI|ciq+pMZT;Z^p? zK?#7$@mY3}dsUQ34_cTJEUk8bcp9nvrA7e(4^)-akE|@{ZB_<)YQhsP)q(#WFKJps z^<%uDl}^cVr+NI>;qVdm87cF7%+jS?;nDeQ^r^Ux+EvYYVccQVNq{W%T1*{G|CBEC z#A&|MWt-^ShBeMh_^WeQu!nP{BZD_g%By`2x{E2A%54~PTa80%Nn9mJs_sB+Ibn(ExV*^hfe{IUZ!yF!BP7O^MeqewN}-G2`Sju2kA zdY`(PABTB^Xmz0PdZ@KL zilwF7sw zi&t6*gsfBF=Im5D!@Ji#+y-I@wZwYAlDN+e0zv%TUwXHW4@!o661Io7+<7CA$X47M z<{9AFX#;fkHYCG8s?LX4kn@UaQx8kRwQw)-GPJwwqG#hzD=rOr24*eKtl#wTD$<=( zA0k-ilw4fZ`#v+HeLeBX&!E@gXyR~3wvr|JsrkA#BDgUAgT zT~!QjpxWwv)~Rs&Pb>4^bx1M)51zBG(!5;|1B&k{_4`=^21Ue7NCL%72zV}L8~Co3 zw*E4ecJjd7eI@JoaGB)> zY{VW@;;K_~Zo`)%z>`|}x7F)MtrgfN@7Fh$*tu|{4mR~>>e5lC$x|6Y(X@NJz)<2D z_Tlb-1>>%!8c7~8Z~4BVEza(2nv~9A>46^(P&(E0HcnGeCQ;eV>J|d-G`9;Vha687 zu|X?PkaB0x9+GplDWsczfAK7yZoEZmoJo5_v6mFrDuWGHD#1|*qku&m!z}8D{>(o- zOT1FIiDI6%g<^VflHS4EA%cc6C()FVkvqNJ5^|E?f9(r9$!-se)Rc`QjANRPsMyP8 zGEy2V(GA>zmmTk`gk|!HQ`yfvnMVhYmcxSUSbOZiT3_uAoT4*p{n`9aOY&dQSy(t& z|HsdgB5S=VfDm@`j_NH#(f_`&4HDB|PmHi+TW-8jNmGq0NE6xYbK6_nG@Bp{%Eq(t z=3y?@7%kgQ+v%nft>07oVy9BtV%^pG%&yPv>T=o=outBQ%L^6y^R5ItZ`4>BhQB;q zF8bE0@x$r!(X;uan?EGqyZ!d8`BpmUortmAhCFif(qFF9wsn28Iq`>AWMNXYvPRiI4#5DB;D42yPCNnZh@{@%L-Yr8}o)oJ~ z!cwAGDwM~=FMf+=T(p=lS~U8Bg3)HzV68N$|^IK3Q*=_nL$qDUv7+=1K8A z8a+UREX_kESy?`6E>Mw7J&@GCg~({8X>uBw8aONB&%eO$H=x&zpZAlmrnvnyh-khW zlg;+pXp^(>b?VP=z*hn2{{OU9{yQ5P>;Dy34gYss4eR}&zOOe%i4|)Ngs85O(dew% zz)u)ygkGYUXL`RWb#KN;o4f7W(cSSlE%8jl=4EiV*H-Q8v$Cav_VsRtT7#U~;Mg)x zRU6E`t;SmCwChOSmSgMNy6^nt@Zn4E2=oK*4R1*_)2De5)N9&En>23UN31L;`E$Pl z_;8bBK3t7oFfO~@z~P4B41b9EWngbpU$KGABfF7a-Dl-&7HzCmprq4z4FsDJ=qZL83!&0SRJ)uoMrlAarw-2TFZZ&;phmu!bmgi(2)X zTIDmshv=}C)<%&kVfN4Iv>n37$MegNX@mVvS7xs7LwhYd1aXK<88o9r78k9mrLBb? z1o3QtsqObgYGBUesWXHjz0k!AdNX~`_o+O}Ap2`Af|4qYUXR;+pz8K0B2y~i z7Ol~%Bo7A%#BNV|wv0b+{niU&P!B$z-ih9kpvA55Yfj+o`WDISH51B&F1tro{lF7H zr4K?>!M`&^6tPnWP%p?_f#f;lGCbBo?|F|$GA}U9d96<@BSnZ3eFxR5^IZG7r*t6L+pchMle`9eC!TCa)5p1w{p_NS*Fz=LHVd;mK6rf*_e_+V> z(p%y31mIBceT;8)Io=W9A;Q65iXrDdFftWkY(@A5&|Glem-j@Y5Lw`PBkf1>i4dC$ zX>G#(Vd{j56%3x)Zb5WYTF}1KoG@s??Rvp*-$wSsgYgT_He%3*aBoMk^~b;$Ad`&8 z?7IOD6`W*43>1Vx#mPybYe_u=0Iqsvq4S}8; zJVWtcO}Z)q2SH ziRTwz79qhkrk|;I!g>;K!#WDP(wBQR2HuM;)HsLK*Y^Uj@zVEjEl56JxfWL3hLd-O z`}&=30)h#>X?!GmZ+-%2M(v&0$?VScqIo^=3wH!gjRohZ5|Rf*`HuU?5RT-7+k^N6 zKjnkXIfQ{32+61_!^@m&X~@R-!qjWxH|^FS-+1GU+}^&mc+uXD?PL$j2DbxO*Czq` z5LL*OA~+b}hX%+>8~oNhxDsh6)>?q~=v$EPJxBM*!t(-sruxYCj{J_p5Ksr!$&M64D_0?jizz&kID7=Rc1>D3#(hL{XnMkook_Rm z#w4EcqA}EPMW7D331SER^5=rpRfkuQH{Sot=ttmOV6rdYnc{RtH+M77oAG6QZ>qnU z;Rk|BTr$Iwed->tpntoL7c|8Kq&rAYyjT6?U0#L(9OY(+upo@7PYV|>o_v4}R3CB^ zX3cDWy>5E=;TCC)?SRgQx(9YQ?q=kLaud#@W!rXsLm@+Ap99MWGx%`M~4J{e9t~pt*0gK_0gi|6e$sL0l*>EDWT?YIC;6eF>^u&E? zmBwMlr$#ss~wwcftLzW$~Gor8b1k9Ob9iN2htBE~B0f@>LWnYp5|BHAjoLy}kC zxAEQmLXZz=;R5}3D}4UgpK>|-RjejLSC|i=$ARge4tpLTf^y>dl+($_lzp4%D-Lqs z%^{u{Wuwh4qZ!UylxCoL61qde-7le;$ORp1NIn^U-18pNdT?>Plw;-_fL~gT!8ju7 z8{$QKWn^b07Zmr_TSb^{5BZ418QXbtZ&m@e71_VqD;$#s@E^?A@f>U);7bz5SQA(RS?_S_SmZUhr$65dPU_lL=N4K)qj)Bh3u=-A zwI}Q{fV~fhHggBEYLT%(aCU--TR}E>=Pgjehq8bl(o%S>Fl806<#|B6Er_onxnM%5 z^PYMy(Ay7n!h7ti{TUe!zzkR50S&J+`X)e9IqcC5+8SDc8iyCgb}`q~7n3{s)6w4vwkyUKAdW!50^f9o z>B9St-DeBr4D{DoJ>e_(FvD$50l$D6JRj%Lm{yCa4Tt2$*`hTIczrR{XUnc&oA!!` zsNY9A0k_oP5$hT;?t zaS9)PR~VtGaE>fXSL6}weSf^R3+rLKYvwe#g7wH27~BQz7gMKUjd_dKl=N1jN2lPY zFGUOIyGufxAhC)+WooTtYPH@n`;^{UETj$m*MCP#-a%MuX$$ocv>0BpPV~&|pub?9 zf&AQUYK^xsPg}zsoF_9*t|w=`GV;Jq7XWy#&d=c6%Ml1neX`RHSo#}gk{TKxDgZY%d&(XrL&LlQn1NCGrTr4DKU2Ov}+~w&ztg1cw z{-wpk<%wN~3Cup8B&=zlLN-s~UT*7$c-3fm&yj-b?D8t>wdgaP_IK=BP`NyYO>#ri z=PG+m95=0cER(hwaTU*r5QOjaeF9AV;RKmw$zg*IFV1@#+-G{V?!T?x0S$=%+nvQ| z%PRyM1O52Xts?>f>eV9<@T2f^x{I5sjiD$SLXYrIiK<=1!oja>H(TN{ElmkXF&;2x z07sm1RQ1@j@PC!zp%YC{}IG zHjc%aC3RnmIDE$*` zUh%Gy+pb^rtxPv-&%^dN$HKD!GKK)i(LNpV)55ZyCXrBm)A%r(4p6c*U2-i z+cY$2h{CL5Nv7qsrM(5VkpJ#pJ#F5OqkD6o@NYgLumKXggFld6vqgU>yJ*IakJ>lC z!P5N(&~GBMd6xEN$g%3f0&uU1N$QdUw-5}#3y=%0;}~!|@swA*N8v_ku4yX3nYSCB z$TR{UYxuvZYNH{Aj6lXVellt!XZWco#wG0fa==i=%_+*Z@@4o_o}3Oy{WW) zo^|2NN1?zG#vfx3;TV5Z7iSD(bh@Iv0cVhh_K^PcPu@0bF72hLU87I)h+cna&RbI3E;E6+#<;<`(ouH`rjGA2sZibI*yi$2iKy)0VAFV$<#s zPGGBQ96NDr#rQ}$YYDTsXB=_1pt2cdU+3^`W(MF}=Q7=zUiP8FLXY~zC{-)JD_4J^ zBFXJZbyiEDLWxx}EgzWr7E#q!Qq!j&Z(WXawO8rX95y(48dlY*?*+Crni(CoP(Hjj zS1lseje6#$;f~y>EB?+Ni5WN4)0LX$RM_8!R4rrm5viV!_P1WXFP4k&vuWdK&KfRr zAoED$7jsr6vW~=nc{QZzjl>AE*>KNT0L)n24heEZ?i~Rxa=>;MiW!etTs!Es#tE@8 zI`Aoa_cY@#3xYq>qAp}KJ5jn}xRe;x726E2RfXlH2InbO(vP?S#rA1O{q~s10(~>s z+6h*t1CJuAtVdm)OVuG4d59@{doM0MqspG&9WLJPw=~O7!?VW7tIK65Wa8EKnaNz6 znqj?@3Xp}ij1l8M^W#`kJB*e zt|^+)UQ~&bJ^sG0aVqJ8PoF5!WyJa5)c)mm+Ul~6fEk0!(v=s%*YZ#(iICTD@S7Gi z<1=h;towuQd4qjVr`ez_hWP=U<5sU!Sb4;0u%M`*dKo=Ti-^$mpb=j$PVv!eRpsvBCL| z==!#(2__g3TgMcaJUS8~Vl_;cJX`=f=aC(jD@W36ZmS|{C2BFcXNrxe9f^77`-!~6 zu8*0DNtKC8jg<~F)_7dATdV(6JME7>Poxs>!k;~bg{ZPtH1rHrvL`5|R=~@r1s8XT z7K?dew$NzO{axbtN2{7jDyDhKniYi`aRt!&C%*lD?$(h|Y*CzTg3ItT+pV*pvH)Rl zL2wB{^$UhSN^D;r;1{GyJ?nM(8B0J7lBLTKy81);Yr7gepoDukXLS1IqqW zWc(t{&zlbAXTKAc-SSglZ!BLstC!UIbn--7Y~BpA+vYnz#%k@Ywp?wLCbs0-p6-<< zW1Vx>{xOqwX_qnIY4?xcPSa_0tkgWd)BDb~^W^imeb+`hj5o(_=-HY~r?aKR=wHC) z9P%Z0UPkBDXex^u6EAz@7L20^bw#4^=Cpj)c<%eOS zWb`k9mrpqL!+vb3QNO%Rzx!NsuRI@wlY|h?DV39i)L)1MC4Zpt-sm&~BsIqX7Of0H zL6S;Cl%nf;7XA-Ed7z-#8m?Qktbi2)5*mP45{Wy_*1N;?w12I?Nz|EDSEjbYM%!85 zmaP4$U+N&c2o~Sqq+HMT=EG31)YMtn)}lu^5p?NzJ=(&7$NG!~j%|%Homg;UmZ@wP zDgA2?@vcekF3YHh4BhEqJUk^6)y~JPMnlv=f%72lk8II)U7K*R+9gq1QB>B*#e8(@ zNgf@vooxdai)dDs%Wxa^Da}qT$A<;hHPJMJwuwS64uOvPUC>1=4goA(cZ$qv-*peqHTcVg87 zV}n>t`NEj1sx!+g;xx00>E#1fRWWm7!K^x=C|$ZT1u6BxIq6^75)&`|-++M$>R~#k z+5qNQ`zEMoln~3fE|5VaWN<(L>VhD7J1Gh%X0(yvW`^#ZrN}sXCqHIwgRfWr1N>h3 z6*W^;UHR_><)kVB2?I#qL}g|rg#WPTXIkdEysIal#Usr#O*;rA0vDf~Aszfw*)`zM z#>JeW?W0bz#bS$sN<0$BHu^d$FRnq68JBG)aPM3UV*NTOpqttW<^!M$H0N2Lu*uDb z^KY+!T|p!Hp#udDjIap`5S_^1I1k=oy!>!Q&Ek>>x(gRh9TC!o6VI_)r`bVKv9tB3 zIq1Vao~!-BdWaU!vA>}wNow?}e1HG8YA~ZwiQu?7BLZTK77c=g6WLWvtK?CHh>&eJ zJGpHRvhz2cw?61~9KWYyH{dQ)`$|&g(3WFrvl0rsR`?&8TxZTYi(h}74f~(Y zT=&}Bo$}@Ie$2%taX4vR8-o5M@EfWCFmre4(-3?E3M`Q_7#E|(|YJ9K_-m0-OEp31@s*5c9! zFsPJg{-z2B)ylc#&c{_Iynr@(X{+d+s zzgOZOnmB>KF(C3q{LOX!t@&ArQ7Kh&W!{RkkSn302pMU_MAU{wW9G}yhNPot^TU0~ z<V3c&e63R2O)z*in={iI+}5<%%BBUX?wnkOKv+s@49$?TyzENYhYIgEQ#wbQdg; zV?inM{mD?JFJJfnFm4pX-<(_alfQ19eo!u&wADL|NTKWCKYAu_(MQa99}Cgmz|$I# z+wj?SRW{%4Y&;ly-_d@KiA7Dn)5(=;b$(kK*&n|1%75P2ET!9d714_e4xh#W{j^F1 z{L+itgk1~4f@z3MlxNOUF zS+~#Ohtb81>nFnIpz(KKl`7Bp$?KUT4+3ZKP_#j=n33@n@dPuB!S{8y0eO`Y#%#cr zpk6K;BUwt#xOvh86Ww)*ilYJ)E$Dg_=qdfOffA5~b}bn=xA~+>(=hYeB#1U6g)AA; zM2j3l7B_bdjVXM}2vu$nj#)hZ3eO|5s>9HoE@sxB|AjzrX?V-)j_&@fFQ+Rn51geJ z&8eH{dE6M!y#DWS>r!jtX4!^@54yK|eMgi^By$F4pymP7jg?(>YxZNCcUL^4H%q5b zsW2JJ09n6H{l*ns4+IN?aG)t%M1j|Oz}d40W^?kf!DC`@V1NZL*uv1bhG_uRH#Cty zqA0{2{$Ozecs&2s-yKm)On<;8oB^jYS`nc>5`BEcF*?NPN3w=ryg?@jQCrX`!TMu4 zP$P3S!9}PV6n<#etWHBeY7yADJgF>uzKjC^g9tp7Uq>Obd(uo_^{bkbwE(D+xFf0} zbx5|!Hj`VZX<%)xHT{W}cHf>-v{jl?y_9Qaw<>`_!7hf`ie94jLJ%wbHsj?pwY^TK zohZ{$++VEHs*PUf$){PvwCvBN#I+e&-^n zvqi|epJ=CF;J?*B*%58l8&SU4G8LGezAEF@^?u))<3p-A$o2y|ROF07W4xo$L;wy} zMdS=icYGJ3xGV~ASumKsRLVG$=$TllVWmegx1Ufqebj(dl2U%etGahUlC zG7u%9+bQ89L@z^I7rbXJpa%s7cv!3km9jmBA+Q!q%%=pK(zGZ}V@UN<#wUj`0ywCE zb-$0~cv@)tkN&h(^xe<*wD#HEU|D@yY>#awKOL`ws_ASt5pzdrqOfxQ@8OfTQt$kZ z%W`YDg9Bm@p&CyEo0p~*c=1xa-MdG2z)*NH_`JYhoo1c7K2OwI{~~yGo}$j~o+T$B zE_ys?i3&5;co^Ft-C*r)<&o|u9j7fofe;!#rU|ex2b#?!{)Al^3L}gmCNOz??ZJ1D zYMt^IC~)D+K6u>Kf?=-oEH1G*qA*|LQv|9;rmc~CJPW4^?$l1b9&`becO5bT#2N=i6LE^u^=kF#rCmz`=%mF(VsUYJ z>*$Pqwvn87oK1J9w-?i*2Hau>l=G8Cj~w3`@RZ`8eEKi#W)-y4MfaV~J^ge=LcNK+ zd$oP@%jBrwJK~2l+pXBxptQ+fsQEFftD`RK>ejEnTh?oz0&$L6hBRx}`0Baj! zQ9AOW^X+oI^}=f5<*iZ!hnKF6q!(`wXG_P`$)lPZ_3Gd21AEq8t97lW<(}Z}nO*B0 zJI}Fikw^0S{68OS#Gj=dy%+qfepnw(4I;L4LsklC@U=%H_Yya8^-{zfD3vT4+;}HdE_rZ-O9X@pmN~D+K<}ka6X@1_$bB3>D=#dl^041bZEob__n^jH;27C z-!ArU_a7hnI(*KwH5$aVdQD}O6-8cJi`%R8y{ue$JiZ?_nXSDJ1J2QfkqnirK_Vmhkp1z z^doeJt4`Rju36cS)-cx9?4T{Tt(t_)ZOtJud+iy@F`S$u=!S_|++nB0v&)Vzs#oPUg~IIY%#Dhz(HT`I@h#=I}P_n-7nny+#_N zds^l#>u!QqADlRV|L(9O>L`$ox{N~6i!gT~eYEM_n}&sO>S$E?v#v9Dx<+%W@ay7c zUc?sPeDDu2eW&~~W1}^OX_WghLFGbI&)mpn-_lLMozTg>&-{aU zQ+}NM?r+)l->m=;f~zooLooqsW(ZWqGk7bvWRVC|_uo()mm9d=u07pnHK9})L`TnV}}1Y(1QVFevOO?Iz_V7kKWK zxQ(_iD$0f}?3xU=^21W=taMo)hw;Pg+11jNHCf(J_M`4Ly?|Eqw@7DxMVAcF58om6 zm_9XuVLuo&;rWLt{RCh>t5=K6-Z8j558s%`myPJeUR@x0;`WWfbbmY=dk6J`xic8!~dcU;k`L>wOAM5-?C z3RW`|A0kTY_!Q&t;6-nSt@Qj?3SsbTDY-tq2GbZaTy35!EaNP48E4s4Eg5Ez%u!js ztC?uf(ms`;bS;}}#usQEwAsGdpYWM@NO%nVN8DCiRG(A6OD|dH>I3xA{>LA|=eqmM zi<GJ_tw;Z_)0=Dd0Nzd^DGy{X=V4Zh5P>JA9w9RgC@aV|K+h z*Tf6`<$PC{#QBpW0X8;FdCW{B${!t_ z%NZ_Y9sTa|9Eo%Bp~yexo3u6_?P|KvKT-X;81I9*2A0uQqVs|pU}r8Ea)&-KVut>> zhR<-Hmsxuep6!^-I_HY7@`RsuEv_4VrRrvf>^`5Ie#g1xjQ&!%kr_D)@axueS8Fb$ z=#h1j^)_&93AcUt<)j8vT1w)5oaUO^I>MAf*)s~g-~4UjWVFyG3mo5= zQk^K^VoN!g*?8C**1z!l$UptwizBjN4A3_;q_R8nulGY!UX8(5NIMk*^#oW6m0(bQQ%4_nBu&?nkvGJE4>`j+8D@{KJtU)+1Fy1-rrxB4kC+jCX&B z;4tH8Zccu2)n)=fgtan}?NyPWky;ayM1vI`A`u}`Ow53%h$Q4`SuxZapR+HghOW{A0a%*>$=Nu+=dJ zYaZ}dpZxvsMBtkD^;PGv>pWGwAg;t*6}?4{=TnHgVX=0GQ;n+!72@Vdr;e+8>i!WvMJTx=)=Vx6TkDrBSD4X?EQe#+wU|M~e3UVP)W{#aG}D_(EiZ83{`$L)3G ze>qr}4=W$7I8Hj9v&Y-ad|e0kZ1K0cUQu#&on{-O>klr*%hzf1D)GsIGCIVKm|Pj$ zF*!xg6jS#ox5;~G@2NpoMt6kbu7KOrdg;-A=k{^`63C43i(B^XUdZqxAEIB*y1uXlCZe)DQ3y!r*CgB(}SN}1u zhe#+cncvS0*CVu83Xf%;J($?4ot^GsWRmGh)a+qYKsz7j{n_05br{*@p~nYHnWN)r z=fnr62FY=NFvuTTwae9LU3HItBs8tI%X87?)*P63Y^eUbi2W3kO|OvdPTj4T5BbV7 zV#K)fMly;>rB<69{zGh)h|oJx{0xN`By13jQaXPvf|Lr);v?+`eNjXM zDAL@y=Dc!Ad&2tRq1Rj+H{z10r68}$>`)tjOm?hSnfx(<^nH)w;KNn2(x{z8QLxseMMXNqN)=fMV?(cJmWU1oovm820LW$_mHn>Q-VivW&ka`!2 zgUNv{RWmhw>MeYy^L3R$oPgC#*h+tZaP%fKO55W>^73IS%chK1O0<_!V@x^v_v5gf zFTw#Xnx<@}4$Ia)YCVx7&4cs~?_qPO^s00N`U>n)832<3{G;1crI8q{#ejkIN_0N_KNakqV~*a(ez==NEMJ3#>t1TEpBJ60~CE^fm0q(F7jaQ<8v)HouoFSYYxPm zplFbatYv2R3po{L zWRJe+2Yo^4VJ(Vci^AY2v;pvbIGdDF6&{Tf2d>}>4@~12a8{1}mvhP^U#J9MDkzhP zn~ZFlO`0K-Oe;baKf5LvEe&YtdP6`$uag8Kk&iL;2086bl;M&xeyKhjE^_EcHpRfa zc)2RAIpsc72+_SJ6(PaYpl0T%Y`J<8nDfe@KT{w`R)O(D#-#NN7{v?0oJL5~#14J* z8OxBioJ=7Fh(HL|i$f~M8t9B^=tAgAeCkQPQAE-dI#RF*6*B|e8^lc9>=j)PH{EYUq zKHpkdtxlu1h!rr!0F zmq#?K#1X+aK?JZM}=Di10Tt_tXfgn%3L;pRv_#w^Il;T>$Efhw{IiKN+STf4st zSpiXba+($tZf+EZLKGeJ4wj9lA%jWoiLvm~-FsRmst|zGtNb(iPp1Hd59ke9r63Xs zNo0&F2uDN^lO&dhTo#%u>?1`f6<8d42uKbgEPx2ZnP0rUeiAw?T%lk8@JCN7@~e3J zTe@VdX}WG&`h$R=39N2ogkFkzL=+tw>H*@hN2+v=wVuNC2d@YW*B4$ihRYkT28`t= zUL!_BhM>knf>No=;Q?iCo5s7Uy29d&`(3H0S;8pN?ojL96wWt@f0Vh$@9jS|3&dY< zecYnmP&qj|wW)yiK(`3zTr^Bc8>CH06G*!*qGgU&4kscgqg#@mc$>R zbwZs4)*nC~fJ&gHF?SQHfBIeO=jgw{5qM1IK8g^&RB;xKTb#=rqK3Zt>o(z^rgM>y zMvi)vMB<_8Q$f`#fbr``C^z&aN@-CiL`$jxg!M%{w_(^UBm3Pw_K9TS(@#Pu#EnEz zC{%le;zW*p3NjVLQzO&%F^94hl*Su z>1jx82m+N5m^?MiYZ4cVNC~(ZkF%1gjF+myde~s%CNjEz<2^A?OCrLPBe}@?%THbg zCU$JeLrlPi%Yp?ZP#cI*5bIACooN&aWpEv#@+_-F8QXr(XrC>u|>bt*%!HN_Nzi9w*{vgD*z zN(#@}KQ6*u+%nXj?R#*5!WERhi%!*oZ59Er4`1Smgl(mnjL0@o69*zb|=n*s5Ad?Mf$8s7Vs=yj~fO zShe@>fa5(H4G}Nob4VV9Uycvf+@S&f zcTWx*U0I%B%^hOGOAJO6{6mH~ShHU|K^V|C0F7J>}^*Um%kpRKad^gqSP?zazk!`X?^fpyg)pymO@ih ztID(;y1L#g2QRac^47ZOq&X;v@3HLz{`U@&Pcm04EfY^&lBSsMI)qA8O?R7p)~;HL zPAJs9#FKnLbclZHY*x-%gh%+iX+*Y^%-^a%EipsH%Cs}69+404OIoL2ul-BbGcUR3 zcjj`z{gD0esVs6v6}ig`9^}U9Ml)qYWjQ%V4h)bgK;jA0t|qyvHB{v_PFBZA8qOUX z?JLbn7lPTj?h^0?AAh^gAT8BR?T;_x)N{$HejrF~OQ}9^{awG8T|4fjA7c-9l}v#0 z&M?xmVzqIYl~*}e+VmXrQg$4Bh<%SfC74c%lK^{1X~B}CNC3Sf(S_eQ{+3A~TGbKj zgyhhL=*p3!5*GhaNj!3qpd=qEH@H^7J9R>Uoh2rbtmmNDg+&XZPK6PIouI2%|AY0` zREFd}k=llY>;)Zc0F^PtGN4gdX4Oa5Tu`Dgo8cz#@W&}&0VwzsmNF}#T}1=PyL7?P z&sse&?*@QQ4N8iK5>V>Fp$D-+u9Wi4JIBpZ_zLcKJ=DpW^+K01X*AN5NU_ZfE;1S)Sw+DbS3O8 zQKHKbx+=ke5=^tjq{Lc8Z($QfDTq%5!7EU_0L2*#QnyU2_y?}|+#e6|ZQjzj6t#i@ z=zpvfU*&P#-``!R9|8wDtDsm-i=R;|7142x#bqgh@~*69Dw<@OgJxS+pzAE_EDym_ z_~ZDnxu0!4AS14!a1B7=Jl`Yd&XhR{KxgEM*d2oRO*&C>3H!p@`Y zj<~X5snYJdvdNOQ*ey^sX@id|3og*R2XtAu&8htEtHVWEBAbWg{-sbfr7F}4OTI5j zpF9nP{*%g-@QeqyGb1U*;CCr4g)iRfiB1K=qJ%VJXcjHZJb_{QbBgPh2X`HowCdI(w zF@d4Q+cLv36N@Q^-E~!gV!4fK7T#8l9IUMzT@J1UMD1HayFidpr0uY38visa zWz(=U8sa(qn8&x}*1>Bh&gO3W;|u>ja>M5DUw`n0uQ%TS4)BWH?{cr@mOyb|Egtjw z{)ydta?j-s9fZaExE^2g^d<@$nu|&(zDpTia}*`vQGZxdl}fI1uJo+*-W+dF-se@X zbv_$D7XO3i58n5~DpyQM#-rItHZ~$IC1=N$#@drzN!=?5yS*jeM(^)kfADa7;`mqL z_Z;ts--C?RXJM6B4i(3FBNe7W?Dwd}5Qtiv7(&H?veM#sU1%y4+lAD&Qiv4W2(N~P z32p)MPO`h*&ABHg3-Mu;#B#DXNghd_NS;bE$ug`cNvxO0m{f#X`I(9BIVpSg+B%uPOcrQ~U7acb6**P=9b94>KqosoF7!~rRp zuos$`t2B)x9*8A;8#|sj9UKH72%yjiHs}kC02KNWR$?M3kuSS}@;}OuiT_2$DmfT? zArBt2f17zx{w%tR=M79ESnIc@ ztmNCOKF&HZKEy+D8N!k7TEP%75R<_V8JCH%|BZ*v(EL~{a1scv%w-K;eu| zuhkQk6W$gO2@#dw7A(ib8aJ?y37V)G4#gs0IXsm|4?7L-GTO^=$O4U0BRo_lTuN6a z6w04CYK=z2;aFkK59PKZgfB5$eu+YqxpPtf{XMx?a-S^PHho<;?txd8Xz?Aued~^_ zxa<2j9DeS`?n&97nR=hl2e^d?7uM9z#hw=-BXCdd`rI2|=Qgo_z4ftNZ|>Oxn>YU$ z*Z%8?uC+t?mLw<)}BQwD7c zZ`*cO%{{8PN#BV>3#~Nvlt@s%PUZPJRp9G_0o6~}$7NePQwi!VxuC3I$TPy0tlL>| zg%Zj-fC|Y;IlsKZ7xG8SL#Z^b@IjLdmZecBECnjlQ07vgEb#w3^LaZ($3~Pwe(;CR z9Vp7islQaDC{5^|J1k@+!31g&=e1)CU6Ch>SYi7An@~_w{|V}ohH6pO+!A=*PE|T! z@Cn}N7_zy73V!wd=bEKV&(O6T`_Y*=4z}44bft=on;T$2G@2$D`s5oDr3ts9{&?RvK0sSDV^>?f&*uyR@0VQ`c?iHr{UPPV8a# zqy>v1ZO9m^iqplJ;wqZ%EoTG1fWM@q9Lg+4kvdj#Nj}LBOMEgTt0ovGlup;pG|Uoa zmds2*-HxAl(*7z>H7ve%O?S6cUX#8ib4}GX)w72g8J(`gs`HeDbb*=?B~neNrPI1O zyhFVs^?P z9m%*JECS(Ct2Y+u^~Qw0)W}9P$_{Gr2EYr7S1_SC%_3vj7tBEU3@u*3Az4V7Uoevo z;FVP7XUrd%87O$^e%8Owmk{WR8MwgzND{x0{Cjc`lpECT%E>q3fI-PXQUaC7CZEIO zp!Rf}QhY!~W^1AYgk|S>y3nO_pfj6FP~e4z=a_G}7tI5-|akoXJ>H8nJ+)<5Mb1^Z6jZ(hSvP($TsNbL+kshMEn#XfREwnBaf$qr+k#RRu&f|p-GNj8bgZeF8EI*lUKG>W9t z1RZvN!bc^fk$waD1m7IrE+6YVM&igm_&F>(203U~B%mGcelW10HJ?&F-%3TE4@9i4 zir(OW&XfTVCtxvJAwfrRgJfM5e1=Yate87AD366#^GkzKYd%jC&8OimlSE2LMjf)U zkD8)7tAFwI7r-u04ZKK=^vCAJGn3R&ft#FrX?nP3*oE^$205!YT9;n~)wT*;b%Uq5 z{Q%b_zU?TI!2KvqWhABAEv)S2>2_}>uBuT@R;@L$kx(R78LEsmgc@SIW9pJvHb$DH zmAdt&2V*b9zKW`9jrp)K1o~WJuv7^iTWLTr1TH%0DiVR1aVhv}|2Rb92Az);M^5Qb z7?x0vP0(m~S)b)$k!XOIc*4_&vbm^;>K{cHsg$@eb^h`pcx9$z>#76^w@bUEUg@;N zO8$U?&VYi>fP&6I&|-lC%@t6ispv1Ok)E^q=sBw^b^cpFuyrVcL)Sbgo?xQ&+pvzZ z{-Kao17o!D%C-B4@xZ=OF|IY!^-@G-io_yeV*pZe=BT~|@^M;Gh?JwaE<$|^oHu}! znjpxwLIqF~-$>)zb&x7QdGaI;+bYp*k6eT}>_Jef;xQR{xY+OTdHj8Y-Ll!PRt0#q8Z06x zbg*&o<^mvKSTIhPl=@9zY9R=cYA!gi;}gj)+qc=z*;)Hp$tK&HY`wN~HZEXmv$eyT zZ`RgFK06W+w~MsO62wQV3o0>Mk+TqaP*#0V{Q-NY()sU&mH7os-~K694fr>mRP1ZC zPFO1@Te2*TCJ0%cw`Rd~eD{botBqUf=)3z(*&*Mzg|2zBK{>D*){;;$6sw}IC;$zC zabd+%75CE#k*bnhERthmf@|U^QhM$1B)o7^PSrGUdTw&WAXY5A@ICeivy1#8sKeAG z!J31HS;m*Az9Skuyd+v=2Ku9+4tQOlS0i~?orUP2 z>K6#nk*!lm@e(n*G?0VP^N=#oY9&=xZ818nSv`C_loWYQ#zU(WVGl{FF8M5Usx?)q zNtzb6S@Wbu6;&npGF@CBx5nKip3-=%qFR-8XQc6}vFb+M1kZF;v$~mY(KhRw-ObYU zilwRr>gBp6?j@dO>DBCN)oS%>?Hb*B{d)Hr&j#|K11cT;MUw7KFQ^#i&e zT7T$#z`etBZ~VU0z0w~3Da}*5r`&ryd%aH;KbbnHKFB|-?QZ` zlGH+JNyTOjJKVF}x59rzIlEB3kYA!<8a0#s6XK02w#73mb*02KshjxOI#`+r<GJ%rw#NCu8@l)ULDZu zbU~)KTz4?fYIKQ}OE}{c2F@$4N@C#SOWsK#&O5Z?O5we4 zJN$;rxbeTU&hPtJ`j2WKOyWD_XhWdo+#g&IvzS6IUYSW{7Bs| zg_E9>fnR1%)W2l@HWRU_R#Oj!o7>~eRuMgBq^{49D zSIY6q@|#n4BsQnmF^Lv^i^ym~u8@6HU45f_OiX1dO5$*Gp}1Szon#9<7-Ze`a)~*b zBxWQgSxyD_`t3ukbsk>G)B$Sj^1_qC$~o!~DZFGXFxCGP16|(2if2;`4Zedd zeE=Xs7wEztMAe&*q;N?vHHrbBhuYao#RZ8KskVA0cM2&ur0|((QEcgOQcAO^A;CZ+ z=4WNAJ{y8}I3(c{nqbdHB|9BIam1bt#_d!)o<3mD7SZ^DPD$Jlnn}=EvgN!b8w^;o z!)d7Al#S(gG)xb>Kz^6V??!!B#a{Z$hv}84fR&s>x!L?O9pGpNt1ep-0c#@gW!l(k z&0m!tp^gCq?|xdlYWBdf8n4ZxQA;1^&XilKC+Ga>$f))SSk8TM{f_zM#-%ZnFtCGln6Pw-0ep{h(Osdmu$4UF*Eo)g-ZHj`fbU z(Y3KHj?FQZi*v!kVacYp#RF24#Bm%vMchVM5wf~rbvzc1N0O2>4$IP&5Eq^8YmPTb zH>z$_-x$BKv|Z}LU8>vFx5v9myQD`-AA=f*UDC_NZx^4I0(Ypo)!lIhtD)vE?-u%_ z0goTWlOB|J5PeQ~!Adv^Wdz{v76JLg5`W4b!B`xQpyc0?imN3)u8u{WoL_+21wX9B z0f*9IA1Z&b59Oq2M>Z+`U{K~ELqVNM7<*RnC-((n6bcqYRUjtC}r39B(5-4q9adu6#CYE2E zt>h$WAOOQkVI@AsQ85aPfa6rx(8ftgSWj(3x}yZWU!NUX4oS^IMZbU^`p?{B^ncMl z;I90Cy}b#19MzdGTBoY_eOGs@mumH{Zgp#yS{7EzmN!X8!nD~otF24jZM{Qv(EOyH?5y;0b|rKRuQY5?0fNlAM>V1ycHW`7O*`uRn(#Hl zBj#tJ9Qv%`q}gS)q5r8RNn=-QNd2ZtZ*YMI(%W0X`B;oKLwVZDM?2ihK6%ipf&&@@ZAhBerqzb!#K zDZmbjm{BR2#SX|28Gl@`nMkdJxS0k>`i?3xW6T;6GE9Ja0W7~2_P~Sh+%9HE4>m=L zZ%;_qPG7lWV}m;pQ1`yA8Gc+d{XksxVt(KjVCj>5_m9&HRgUQgp$s8IEJ<)mf>jcl z)gs%bTC^CPjX$+3a3Oe>NK-`Rc>5()bb8?m;-4?MbY40r9RaTL71ENDU7y^+HSju$ zG)CN&+FV=*phR%hsWdK_8kE`7A+BujP!$~iin9h9Xc)O^@HFtc4FxSvwKZG_CF^ks z3~oZIbhft9RMewUQ!yB%*p5@Clb2p7hNu#tI1YH3mzR2Z4X-U8(za(+0Iupd54vqV zWE?c{p2=UNf%QV>4D>tC@S@Dr)JvJnY1<2DNG-3jIAjcbtU+`vZ)OfTTo~>6s{VvQ za%3F)h5Ze82zMLrZqfuDo;F*@XopH4=vDV>7RU>#dfSQ}2ZK6;MJo$cOIVL#iE&9& zPs{waC0DFAZZqDl|ES@kM(YaCZ5}b)v7tj8)HMi=?YVSKo1pGEZ(YsZ8x&5(=k{g+A{F?~qr@dk?@u=0ljU=#6XJi|ry_dzheO zYMSufPD%@lc9k(VRl{lq<3fYkWNK{0hkpmZfM#XG(`*}%uHtk@=o7`m@zMBLTvd!; zh>N3f7PnEQ_|xKiz=66!l7|aG!rE30RptwI+M=Z(Lk2Obg}EBLpk-tLK7a1~oeH~j z*T_hQ?CCRLr;ywa9I>*-6x1w)jyEqt^Y8qwj`O5pm9ZD7P9~FAX0b`h1o`>E`4Wg8 z=bJ8SniGhsoh_}+ts*d33_1X0RmoMNrpZ`<3@PXg2o9$;Y!0xh=p1z+AhhZlW!BVa zbOdYxW~oBdrfCldTw*~i)DnNDx|)o(Fh^pp1nX2FI-|qESLlquhhZJwG)F{%jVC88 zg%;UD>pjhxGSQ|+&2rc=F zkU`_Fm~E{K%Q0?cL3S5zMPxw)0LKRq7kxa|e8q+jhSEQLef9E=ctXr4;`xd3TdtfF za2P7AHj}%3_}aQQ_L*#7*Q(aukKSnaedPN2bzL7=6+3usRaLevSKnB(YAhX|m$`lV z`?t+;Y0d4e_jldL*0=kzgN4N$KSsRtHn0Lt0gufCOR=9QWB+JKO=H+bqgm|&=p|RT z@S2F#M?>929gT{3hDN%H>YJ&)$>gmNRHEKVdjq?xs7EyyB#CjoF){!q6!F$OUdkxU zbv{y|g>8Ed!xmeF5~Yph9M~FqC$6Jw=-QB4oq%OLCKAm`uSoru9_P=;=@^+mJV`|+ zQzCBXQv!zm6y?s8?srBpv#u=evnAMzn1oUg;*)9nzhuN2SMg z+C^H{rgJ6Doz9S}%Uc0_k4ldKM%m1h)P>b!YH?6Kimj2Pe$C{8t$xg8vO!mI*gR%d zjUt*b19FQ*w6(}ina`TFW{m$Q+ndaT@$dAM?A=I%SkW^i) zWYi{PDI_z$!5a{KUZW|X!%(syhC`n5%#A^or;xL5YhiGQ!~B_XgqPN zUON~WhOiQ!gc}A#pdj+XaVNOM$o$|0ttr7e36ne~@XY#-)8~0dIYXA>`ent3=THgXc7^}Q10~?}) z-oe04-jTq^9QQiM>|@?XA=Et*JqB+c-_?HC_p0t(;P>*I%nKkr$7;vj;k)J0=!K}( zF0*d}K3zr#W6BVM0!_iS=of=f6&Hnw4ZKQH79Wm`&C*c25HUxt4ZaL<-*-K6y*5a@ z2v>oUVymNomo!Fx6gIIw(;<^+%G1a9-!6c`sD{Q{NhMXDrIN~s&A5vZ&l}KR zLPLES_M`+y$fq*QZc=yrXSY54JI0RR`FC~MIU&0-8lAi4imSeGaMP77jqK_t{+?-G z{u#3zTAIiw+&6_o-J8Df<+tbO_HjLnu1n`2KfM=1pBk~J)VmUSZYPk|czL(dDP4?r zH$e`1_>|=_%3J_JJxb;R2$|PL>7R>=biqqk$bqNGu!I(CGSCV;IAg;JNM}eFTmtx> z+H6+hR0qvnc?4FXGF}3WF~z$lnUa*g?lK3ti;|6!M3)AJgRB@Fgz8h+h-Qo)K8tx& zG-crlU6jf&NkA;p8QRvTc^&?8ggl+k@p+b`q+Vda?N3 z`J>(Uzw+n5-`9sZ(w{LidyO^XsjM+hU&v|NH|N(ZSU0g_-L{2Sy!{XVz!ok2+UNLG z`Swd+SQM~FcYUAztZTT?cis1X@Xx$Y6JFJFV8ggz(6qnQr&Br));trL%L^8slxzpx zwLo?xH6c*|2!}{Y;e5WNOwmr7mGCrH8MkZ6>Qp3klH+u>md~)L2XOaGFY+sp`N5Mk z9#nNkBcB|Ij^>tYBs^QsJ2sZ&{F$lC3v$ps3INmzK`QH8^M_FGs94L!s1w_?vUWlX zkoiF^LBXlCpHzKOH3_Hy+)9gAqp_Hv4()P1~l<7mFeo;fCJ5jAsB5 zf@fWS8Um4e1sw<-pNkC~8@%g%gTkQeMM>?G10dD{1rO9O!yKMqes`lTOv^-=cd}F? zFJ4}qYpm4x^lO~gcs2mfbDduc)Ig0E_@rw063s#J9?hMmkJ)Yyeo1`X`-Jla@#ofG z*xnTXLvlI*&Y&B{+a1(@Tl+oh1ua%Pt@)#(q^FUh!AR+D)-Mzn>HESf#1;BY@P>5I zdC>QO^UL}#8&2w;&`%h?EB>4Kis?;*OLtZa4B@j{aTlfZ?MUfo0_@@aDwj~}anqwZ zp(nb*eZW2Lei?EmcjZ43(gKu3U0en=((J!L?5md`63sw&s@*MCS zf~4$Cmur-apvH7!Eo^#T)=4%P_ux%*6EKj{=pM7URl-5)S5mgD{!u&PIH3IaAcMb3@0_UQQaDIjscQZx{reaOOj6mS(lPO|mP(rf=^l7z9 zqMLPOP>u}Bk)9tFE&75RszN^5UNFmCot9Tp*8MIsEvw;U_UcuyRL*h*`i@OOPcAp?xn**)>RG~pcV%o z`ic^!<|)O_v1ri6;vslMf=R$hRypa=5Q@(EDce-X8E9iE%hD8))TeFF(cfuql~x+a zr+{2Z*ML=0Sq!E%F|?#hGd>WRhDIRcv^9T>pB0UKH1cvB|AN!l7E4XGya>HS5pI9x zjH0hwS)6;f{DAv`gw!Q=3rS zC&jlWS#{((woh8EEJmI6*x+CzPutp*w4*~wV|6?&dIFKgF$<$Ez5y0W zQ9w~5$g|@j_P6ct(B*UBA8SDROz|TQPyc5U4>U^vFjyzvRDIKOQK;7BG|TRmRUIvT>xNqP zCWhA$p5JX1AN3w;`q#ZR)&Kd~k=8X=X0Ulq+%*D($S$7caNR=S3#xFmwQMekTQ>n6xbf%$CyNvW=A@^u^XKGyFW1+MZl;!l7BtCyC- zfVi$`&}e<|m6y$0mwrpC#d;S)To9H@wZ$;BREBB~)sELnwO(3ZYfUsHYMTlGB+BT zKNKA_VWjrT>uT2(*B!<;SFfW>14dKhI?JK^7qW$%m$R@=_Askw*z-Il`;(X6DB8Jh zlZgle4}V5zI6YHzJ^+~ztor*Uv0q}Oq7PVGd`e%ik*8=?qT3UpOqnO?9i{DS*F6P^ zU4$8q9wfn#V*WY;4iC?q&y1Y2Wp*jno*4zEXP509Hj}#uLB>r+7B_n0td5Lv?8C(t{?v|d_Z3|*efnbG~fjLlL z*HG73CuuqpeTiJWI{WVJi!m`W?nRR&=2H48v z>Hu4{I@nf;>XmbZ-nwp?^>jBi!@Z77`fIB80J}1ORX|vtzDgDrRLl=32#{s1y;78U zmr$shLqpQw=f--w_|jH1+6|4F0&{n7>1yon-L#-#`p@~=wk^*bf4rgb!GB>_rtcfL>vU1A(O38l zYR%%|(G#vj!DW}VDoL$34=>xb`ID>bo4wxnJpJZyT{wD;c;`*Ge17%3kz2;s&U@#> z4Qt}HvAG8pH+noOEc!ri7Nq~cey&+OR9YK@twkERHiO;37ea$KM#o;Rp64Y_4~+>g zapm|P&ugJh>P^sU8K#0nq_HUp#dQEUgQ39Nh}3wgjT#d4P$@ERakf=Wd5j_RdeO=k zW!_#5kgTkofUFZ72yuiILQ-fXHI^pM%W7&ClJ;O$MVu@svS263aRrFkuj~jp(P}&W zTs;&El#(;pBA%YLJ6zM~pds1BW85*>h(@VNd(yy{Ujtuq4SZQPl(=2IsKo#BwzjYc zFOKly2rrJ{Z7%TH2AMaG(+OnWd6Lf6)U>vgR&Bm$m-1(b=)zkl95N2tNWAiVD;(Gu zS_eTlTjSQmXzN(3YNBRML4I{{%shYDjs+z%;M-Ol?I|UelDA3se3FpbiyioEKP+FW_ z{(u@v4waCd_ySNC1H2+F&n0ddXAcu{3h^@R&5DJ2$)ZzPtGV{LqD{*;hh%rFpy8DE zjBM%9M>?+@u5?-qwZ-YV?qa<`3U}4kUEl34ESzq;BI@#5!+v+(!W`;*FK)VJ!K$l^ zk4=AbHF#}gGhw@ub=`kWzOipQa7`{8i#ZLgtE4NG{Uhxa;cHsdV!RlIDzPM=IVHrf z5Cl1JiJ5nIbAeBzD(fjYu~IV+HtFPI|q! ziUGP3SWZzh9}h5zLAN*a@j!2;j0Z0bd_I0q@}x3^V$oS+K-pvB*?v1+UsW8J+wR9}e6J3ylFRz2_A&6)P{{JHk|BC zziw`xaixKNovxlUr@Fd*&WC(;ozwH@=b(-d@&{5D=2G8FC)%s4=S)W~%Bu?K$o}?~ zY~!c1vds*o^)AG`|Fy*ZC#L{E}Z1rc3m6Nm$UNlhk8PsUk7-QfW`+k8NPzogvPh?Zq>t zj+x1nyJqG&>|Rbh?1x7>mZ5Lm(b8njOkx5oX8JQ@nXg*D8a$lQ$T%3yNH&z5%}9P- zDk*m+L#Zwwy^?07)32}gRmy3T7O3|Y$U_9dq{W@B;~ZLcTd}=b`Jo~8l5!dNv%)W| z^jpsUqc=qy3x~&KW|bLv@VEf|9GPH_Kd_IoUG3LV}yQnZirWsBPc4*+?~~% z`2B6%_N${q8EF%mGnL1Ak2^myKu~Ta9|Vp}$*GEh#t{fvtigD|8V<0KrIP3zR^GvI zKUeSVGP6Wko`(1IoK&X0J%bK1`n|*J*3`k1fPG`coAX@mKZp3~>P-8z{LZznotqbp z)|<7fDf{b%J?}6;N!8P9&T4G7Zp`;lbQN}hhoqN zzvI1xGR2HSC^RQ?VOAm$I1)@m!-<9*rvY*v>R!1rt&l6z3OSlhsev2;#c)&PoQ>JT zs)PZk=_!6ABHNz^#gGu1(EA+CoMHg*Lc(V!St@ci)jI5fT?6 zw4S;w7%Q#S&tc{yB}HgUl}&*wi>IuGa6xovYz+VJ(|_79W*jp;WPQ;7kmJGdc;T4A zQ1BJ}8*Cfw8^SwmL-wKYL!$n*(D|@9s{gR%Iq5m;tKzHH^Y%YFbRG5%Z%4RQ?kFs@ zju`e@b$PMcCdcJOz5t1UP3yL;WXo(TWL4C*nyt3}*7hfxdWn5;_&fUV82;6uuF!jI z;b1ttK%8gQ80}W4*>4J3Lzb{+xwH~enE~4h`wFMVXAK5J;pJjkn=aq%<-MBOBtsHf zQNpcokn%x{4g>6CO(xvBMB#A8U?fMfM>;0Lna}0O{FTq;xm=<3GHt_28=`5>U=f-T z{9(x{uC!U12u~MIn=kAS`Enpvl2ry#A2JZNl8iPd^PNqh<}N@v8!_|52z*k)NQm{WWT%G&TQ&?-x>aCm&0Uo1tG%_b8~(^{m&nao9NzIa9ihRZLY}fd2%}GF#Y4X|b&^r;$}mnzuCfSFmD5f5jj$r^YG{qoUT74r#lq zSURLhM#~~|Jd##p5LnD?DmA01zBJ8g95b@uCp%9qpIN z6>OY27mZUZ@1{`7edB|^#zoT|9XY=x?DeMrgRQ>z?ao!LK~9ZG#fKItG&EmFG_~M` zt(N|&w2o8^@IB`M=fqan`A=(K*;`wl@oVV^rJt{sD+bbUT$|ywGIcEr%e5AkYc1sH zSyXFb=b}ztuakF_PTo;EdtG$5-tu+5H4e$?5E_xSyjifDCB3o#R@xJBVA@XQl$`QX479iviYA^ln3BMN> zb>3Fx-#}$C(5O8-@JEj{iY`7tdu~V;#|E&qmeV1CP?3hbywNb zp$hD-k4H1<5LCg;$D^I;&_ZO2bE7MhOFLJDR(9#qElZ1qbV?^^iN#`=L*4h>}+%nN3wosX8>FUl{Z*O>M z|59;u>DW?HSZZ4;E+yq^_=sIPxMtv_xE8C{0q;q+<#tY%&SU`|+wSdIvMb^-8=9|}Zg;9ek9BDo%Ib>2Y|FHJZF40+CFiVU*L)^>nI<-YfA+sH zmo_aiT8R&|%5D;UvG?nP<`2MV} z-b^kd2;8$u;ky;}F>9<|>(440WAFyS5j?rxSmE8|@vIrdMpKAQVYXeKIIhWiIG3LD z!RvUQKR}(*zBQ)f>!*&-6*@j|=sa8z!^4X_Ix6ybpxzSXHG{k$$WI3Oi*j9!@`RD5 z>4a*e(@l+kE7J}1X95-oYbdv?H^wXtsA z^c-ig!Jz6~dh<<23k^lr!a8u@N+a?7cY0Q}-#Wc-b=b#wZ`M9`^Zs2QnGUY^1ToeY zZDA`OS?s5gCSux#ZY0K;RS1fv(nt${5L6T$IYFoF>ujVNZSt#V?9fTdz~wgzRlJ(2 z0@~BIxS*_!iaMLZd$_D*p{Mgy&rcUuQlIjxT)fA*Og2u#**F2G;*5_M}N@z^Ga>FSl^`L zN)qE?_MH3sez6#4OLXO@ad1Zu8SDZ2&=pneEXDn*LDjHoOf{j>s9q;0Mjb`dIJP-6 z`T?XRObpsg&qQi@&qV)KYmb^jOOA$BOO~x&^9<=l!$;c`bm^vO5uCr~n}T1ehdIAX zs(;n?YUQj$tPca1@t~Sn&;ehTB2a7$aZSQ*b;&}I`DNzOBj?p3=QP`7Rw*HJ8!KeN zr$+QHOeg}a86h}@40;{(5Q_G_;$F=yhFdJRI6mOn>)jjB!PZDYTF?h<_Ch6sn*=sS zMPNg$vy#&X#laU)rAYGPC<$If;p~TQxas)=&)>3b|Bse8-7s(bBO5=oeUWtJ^LHNk z;5(y_-2IIY{pIG)j?drvz3E>a{`)uY86;lWU#7dIr_k4uLP4x5^|kaIu6!l)3V~Ofrr##m47dc}62wmQopq%A)Z>%QqhHuE+*FeyY$>n3H6uWd4f7(y@C29S5`9^=CtAIc)xY> zInpUBYXIkA{Q-cFM6SuGmC{&}fB>}MVT|-N^Z;C6*QU&jY(@*{Y)K6$L%sS#hB?$* z3$`V;b+&_c)g4(jC)+WnC%Z0tz5V*^ZrwinzU)VJk7!@j{Y7uCowKH4pm9f|syK(` zby6zraDuS%-BINPd6kR`$w*%^By@?6OiEJa0G~roB5LWGd|pd^UDz;Y5C;vTh9d^a z@R}@gMqZ^X_mc`M#MGqvsi=FZBZF!s+qmPSFPwkqfZJL#{pvLhAp-L>)@_kC4b*xq2c{nzingt}^4e2B*>$DY!Syw# zyHx`Eau5u}7fOBEpWstv!p%X0-PlZrO1o(b?Mk?gW(jMI(;*gO1#>!v(D1ktjQS~M z^IWuPgBokd5siB_=>dmPYf<_Qx=hFH?ei4xu}Y_zs;s6dw6dD|VP!S-LBG`>T4j@9 zK|w^f>`nI9h((xKd^vTvMpf&tjdWDEX8LSJe=*Wmy*RVR+V39-^+(o1Yh}o`$-gNw zRDG*$m;XR$SL8tEcK^pS4_oi|KODM0^4aRoXTIur)c<(k8<|s{Z{Z1kk$FAycBWdc z*&W}VI^_I}^E0lg8tro6okR5}q)nFg(3M_mC@e+&X-2OVjl*}eR%5BGgwdLXx?Nrf z!=1Z02tddq@G!39JO0-abv8f}h|jv8cmEL%=($FKdw#a8MM?NJFxD7R+y~_O7W;h1 zMe2`^G7hshmU33aDiTnfbRvp-qB2XmNSU(C0VciRT_dd-QVq^r-i(*LLX`9<1x|st zNcOTbJ>m*WH+UPmr|X@qL6>*kT}y8NN9Owb!eFAU>9*vSj^V>!+CArL>B!sHuBi{i z<2GXf#P5#2H-7jki_3B#b}`TX4(s)|zI|${L9rQw>{xsfeJjPDC~Hkt^I1|8uCOOL zd7AWwS!p{zd-oeIQ?}tUO&X^8i0ev(`Q|*#iP$jT_410~P69-kSh`+-+Wop{=}QhJ z4i_ZzR9%){!pP?dQ=LUvDqjI|{RePjH+>J2iK6SxdaU^QxE; zslJQt*SmXP%i_jZbhX>zuBmmJ=gpnYEUfYw)No1>P8yh7I`ZQm&(9{C7r4^bOfTt8 zf_#g4_?~L>;VS}sv&4igm(GblMSrPNHJ18IvVr%Ph9c2tBI8n1#-*mr3INljiOQ1^ zYk2~)(o$N_5vS|4x}+7ha~U<;r)E2VQytF(=BirjyE()*!|z&L_On4iE{c8!v}>WS z3Mwd%6v=v0)21#0%J0kzXKXJhYw1kysE=5aI#snNN>5`jL*@dW;dums20`n zYHe4DZ3*pxa$FpE%$ha*=s7P7E!^1230YZDLIJ{40Nv@mu`4>k-pj5cVj4JVmgyb}}V z56wTcoQ|E2|8w-k*w3@SRsA;l+t{mFqoX6cKD)DKfA$bNBp#AR-Q<0CG;pxyP|i%K zX9hTz(gY0I?^S(2stZUSmjhk_ed)^V1NsLH56SmM?~5578FMPzo$YJb(C~ru2eNlq zz8XE!@H^?Z0aIF67ZRQkLoCel+!VZ<+8gpcQyKDy{mkZ<@#W~G?->uB z1b7f6#731h$rE*meOJiktLtFqjF0O75V-ba&f)Up>BEXYaxmr~rKCTS&Y9#Y8i&c_ z_po&g9=}h*zYx5d_;O+Bm1W11Y>*oQOJo%+CZB?gpq|M`d&+}@u*LZH=pyOEUW$MZ zUdV&8oV>)4CFQ7dZ*rHX9E^Mx=osiy19oaLv(aTX8_N)60}2Ki81=JJ$dE1B3Fl)0XVh>j^YBOKr-tfsEKC^elj)=R$9Ek)z zziGkhjnluXsd?yw%}X2XHit<%GX1^#uJ5jCO{H^-Hh*#d10jQ-ExPC4s|pLQ8EY%7 z-u2lEtHp~s%ysGC#dg)Vfqr|jtVf@-J2K+^Ch6mAjFaq)beQ;a4+x31g9^t7#M+wPvWZQHiLwykN~wr$(iwDGl#J%9F2&Q8wB zsbnQ}vsNW_QAyRR_kAwk9+lZ8ZdOy9ac>$2kVc}RPg&8~>ktdpP<;Bn{DGR!jX6m@ zyRS`$f^KujNtOKfP>WFaxm|mrZQYXov7O!gy+f2vssv_uO?}$9 z^}ZMhQ!}sTM8~=vt*H1Fn7Ye~YCT?^`DuCF0-=selrz~mY}RP?FmG-2v(_r+lt^B1 zF^lD49qU_`YI!AR#FtVeoC~w4$04QA67Z<6m}jXrup4sKA9&gj1{ALyV($QG=^X8F z38ZQ}0Ie>^7{wM3w*D$eGh&khp-uU0L4P+W%S0NJj|)?>7@zeSD-5z`t;#e83C4re z7(IBE;rWV%e|BG(xM-4CsWIaKddz!gl1cl$F7z@)b`Bt_jsS1aKT`gJVibGZgc(NfSAWsT%hUs|Zbev4>_L#il7L4W!|8$6gy z>XGFmJbnR>w$B-+QqD+=0V=AteyLI6R%xG;zD_E(J733yn}i9wPflKLzE^~5u?eCS z&`LgoM`-^bbxU}ti5gAL9Vbn^QVq!zZxTulTjio9fvaC9DeQwn)LBh3-}%QS+L*N4 zDTqz9X4&BQ>A2o-*2VI=A7`AcF>f=cT>;)=>)_dsDYSsw{Q0kAJwQN2PVlv( zrL|Xrx$C*`E(lVcW9;+|uB7%4eMO#`UYC&(+sjC=QKvy1AUENbC#r?FGJ3Tz?zSv_ z?j+%|IyQ;hI?DY7|N3qT1J(xKUx)nZ-D>RuAT#X^{SDo%{VHFT_wR46@1T@JRejW+ z3U85b(OpbXvL`w$m~6(Z8~!uziEo!9m1X&Ed+eRNo~v5y)kf+Xd-k=lMVaix&;=>D zD1jXEBq^L2>hkS>y_b1Rizc;gIKoHLgSt1h*TM76tFVd(+YT{dQd-{JMHx}R@#>B# z=V|vMMKdRcw`l1qx)WTAlY|4-C2@2h4+#13^SEAaty=$L*~7cB7rqlWU6)0r3(-Ut z(hs@75g>~YQZS@d5GcqM)2UkyW0!#f?$ZTLIqGI)#)siH)pi_{*Ykz1_a{UO5xf^~ z?ecsX167f;Zr~faR1PZ}>>HI$8pw?!7^g&`>cr}AIw^UL>=={`E#E!A#9rDAh7Zi{ zUNTM$VCL6-a;(SftbhIQS5W7RY(H@FT(J9QYpA2Jj-J5b=U_po132>9UA_nvLYBnH z9hy94grWi)2MiZx81&#o3n%O#DV@cqJrntgDxqW-mq19x;pC2eXZ!cPfo{y zauMCI-BJ>_BhURa%4_fzu!njY99StUjq4c_ztoiv$5=c|*sRLXi*68YNl`UH4nzM=*; zJO~~VZ|d#@^#sj#9=MZZm-O$A#0ka-kAg2XC8dcYX*|~6F z$A!3Ye^xq%|s(}Km}dn9vrv;UE6UpPa16Xw;dT4wdbq z4Obc4E8TlE#wytnd@JhKU5xAaY?qZ*mCe$02+t+ zUSjXw_oOlbeVPA^g4FsbwTuMF4;+wfCcG?8K`nD|jI@Dxe}El7_ijOFG1c_Ml2JimLOSkX*UZ@i;Xvy(9sRH9*cNU`?6QdBCkv-$elK4z(kO;szY{_qgD*qi>P z?DK}%zv4DUS)o9b)%^mz#C2~`*MPx9P_j&YKft}bGF!*~dFXimdGnF+> z+9Nk!;2ZZiP3#-gs^GTuapSl4Oh=1m%7uda8uw*& zM661FeT!{G{p^alK9*MEsq7;w#-`e<+VK14lNB-+@*2__vImlij~0&|ZmZw$y9FDat?oh^(mF-sP@<_%(O-V`H zc#_Ntdzua+BKgOYUUzu`G4?aI_C0K^ATyg)W>uXj++ph2$*FlCiCydBM3!>DLjyS) z9M0Yk_kL7ig5&Z2y%{H&Cnmz_DspY*d5(SWVfX&h0x+!vU-_u%o4GR6P4Y}9$tjr8 zy?+>a%iGPmNNdzof~YL6vfqw9k~md;%3*{D3#Z!pdXD44yCarzv0-Da9I?;+C?+6n z+qus@NTu2jF|)VGvixf+g+n9Awyiy}S!(9it-qJ&S9-o5ET!Zw%O5RtGiRsl3L6fqXjKnRvRA`=f{zH3kMCp2BC-GRRjCHHoSS$1c9;6*uWpzaAV zpA73URnh)$%9yANs=XO>7k7wJTFB||5BZnbtawd|7ZDY+FiNB&DOQ9fuE~weJln(# z&G0ZfI*|*N2GTQ$rPH%ZO`Q~jY1jIWl=s>+5oUm{AbtYC7EwlwR8=hNB3=Y z-GFi`h5hPQRbBUOvtFL?h;ePUyOgt}_wyo+Dbw;Y4ZpVs?oJM>5hE-0n`3FeuOEi)kgv_iqcfvheSYcECOj@I%c)W z9wUcNUbs@fe6 zU-$cj%HxNYB{4WEVGG47=Ql~~A@fR$-wbk4N$@2X>-vjSVLLGTo7DD03%4^fo51r4 zH4cv|XXue(<9(ea6P6TTQYoG}zwJxw;V$9QyNVIj!(+$ERTT==jaxXgnluXG+ zSeHRXiI8E-Zn5U6w}$g7RqgXsDsokiUzHY$nMAxVc^QwD(8K~|Y?d8WHgPGobx4|` z*3}jutFWkKarE98&>Gcx9+PGlE@9#Q=rk*p%cF?Ig>dojOp-^5zfw%lRMpHHRMO7q zD*$7HR?`Ym>mW!1E=PzfrNixkkc`bSd ze2`r*e)(a{WK#%kp!Du9AJIEQh$7Ltp>cR_%}r$$z(^t=`*xR>vQZg^Ws-vlS~HU>QSL}j3HQ%b|8Uh}o#an!pUJXN{A7F{P!+`5%!Xc` zeX*5^Mqr5F5#UQIFXohE{!e%&IF}HDSMzOlZbZ-`gN=>5T7?phWZMf-r<1pQ$n==c z;x|T8h41{^_1Oy#gGl2Y`~CtMWSFAv4^*Ml-!hsl;y;u`bAKT*;OMfpNM!i;o;F4m ziRPn6u@wAr%l_FL@D+&d_4)5SPQMV~a_(+ZM-#f~rPw%rnngbe4Sv>JSsHrZV)T~j zQsYM8sotXU4;FA0noDC!D^9X`W$K zxn0*(5+j_dssG~X5KsKWG2xwKhfp?iA3B(@0F}Z?Fz~0*bt1xkO86iWK%v{6S|Uk6AXp(?C5&YAk7pL=%@$Rb*s^tYr8JJBXMxj^p-H zQSnI~M^wHzPfcl_@k;=#2PH$xYHS+S-_fjrG7mN z{&{mRYCe_fLKda%Af=_1a$jcHSpl;!r^X`(-OS%aTI|++!=;pts0Jo0%Q`tem5eJL zE;P$7W+r8hJkZWVQWgv=A|wkmy_hv(lA7$-XFw&kf;4zlv5>Zsr4-1bT0K35WOWPe z7m&bN&U^adJsF2V7SU_*L11R1~+o@}}#%sSQo_Kls5QCVBVY9=44xIFP znknSa{csQ-1y8?iAFb_rjRB9*C)99wS*j*9E?ji3)pL01gaC#Bc=R{O|0(YJeg$GE z`AXIaA3$+>pu8|H|9o6}oai~+Tu^(0`k;Ib=`m6O4&oHGoaqk0I?t0q&KWM-u0%Qv zQ))8Sog1_3o9<=C8;S?#*3kqjo#^^~-5Dum04+ov*xB?~;@dARGhD2{>gcXHGHAuZ zc5^atYwR-p2K(?Vt-m5mL{PUxEFQ1e9s`DI&^c^r@63Z`6nRDBI{P=iC}`J{(R*HM z`+*8?AGgZ--rlpr-Zp_Vfx>*XwOyHcF)|E4)HZ`#pFbW=GZ~#rK7hd`eP9R1b4cDY zZTfGVK57EwS>E#apcut`cq_L|hLA!vI!JXT*_=>YVK=b#L=x13{-LoBpr95N15Oaw zhNuHL!Fs{kC^bUcu)*HuyWr0~8A^}tkeOa73L$ebJ}Qs3Ld{mKKLi0x_=f;QC51Av z9=MZ*4We_v^TNMg?z!r%SEa+Yogv!E$-!Fq-W))uOvFmmpp$$2h97rL7R>b3eVeC zHLtYq?wY$wi~)`|ecw2O*>Drfb$eyqmKD! zyee{cnZEb+E&Wr^9wyDvL@&G#A17W$;>kpB+y}81F7mJL@|`8zBZ0bX8*J&87Bj}@ zppDY=9ZQ|92PtXq3pU%_jQa!{!0N5TEycaoiT>teen>4XJ6(Ipdv$SnPv)C`*LleI z6L7k6N5|})@O%0*_nQv*o%z+W2ZYGg%|YAgQ~$MFWu=DCb|?QYKN<*3 zz0)$xe8Vj6lDny>r965Z@R2pA`iC`OBNUyKA=}p@zY@6P^(D@HBKt>fxx43`?Jas@ z;H@p}lAFDMTi{T+oA_T*q3N1zU-mcU{|X7u!KxFEJM;V7E=#rckh>1nS=51C(bb`Q zxxe@wue`tv)!pe$&GGvj%_Dxmm0C{iUstz#9|yJUn<}5^)!vwsO3yD;Uj4!Q!{pUz zFTi${ce+_h<^=GLjd9=t63pI8%r3+syAJ(K$xvb~qE37*Y&}ux_|O8vGm@jgho?4S zFkQ|N`pd)+d!l9kjc>pX(1d=bw$uZAi?O5=aL;k4q}lfKY`SaZ#{_!;Sw*kxTGSu* zgN)~$-m7ni=8v3b6$Lyco5W4a4thGsT9;)qetw481S>)2KP+v6pj52gQpJ&@U6_Z@Hxz zy`943XVspQ=B{DBkIz$&wBE&-51QoOL2`iYD=uB1`nhPXY7Kt6t6AD}I9=PM*-pWo zt^M)yM)b=kuo6#sd)i~S`-86Krb9Tq_2z!COZ__Y6+PM8mZ7<>ed*8L6KlcW?}uaI z59{m0Zg(E9$|>*Jl63FxUFWM0dE4(R{WGTvZ8N|L$__`eUV~=|a?kgs{&gx4P{Y;O z1KJ&NNALaeInkZDp}yk^FJ*UZ&p#wS_1;^4SIp6KQE5l_c=!C|`c~nsV9(H-nDbPU zo>SQ*%;ugq;`fg0D_c`rZTm-Er&U#tyUE#Nwr9Wk?jt{8KR+SHy9N1bYCN0AZEUi| zzSC~j&U1ZbTU*?2ql2u~=siZ;eNx@7TT_qSvc+z*)A&&0^{wSO>yPxwrkXAO zdJOwfiH>b{mR5)F{n*@fzV&mFdFly{Kp(^C^2fo~k+gbOYxh;>QOsRVLX39HPMhOj zD9_W4@4MThE1n}h&Q_`dPbrb+2FLRSA@R;wR zi9^5D_ujR!Mzq8Cv<<(ptuu7YbseIb@MYDVj`0m@x zF84F|!no$UXvp1)o+>_<>9zb#?z1;Le0P7UIp#-2voBh%u>iBUhPT)ULEfZ<>{u&P z(RSW^+}D*pQVh8?@(bF@P)^e8fM@jddN3Bta5=reQaGjjzbWxO*0p7Pt-!maFr9$-61$xCB{4;yZi zTbPr9bcv*o2Dzp}l#(QmkjQMob^!@cS}1Xmuv1p_qVrBcs`!O_C7`xnDG)8+L6{2= z=-C4Jw11lLQf+J1I9Do_iglp7+KFasE1szNq-a@kT(@REIfvWAG>9a>(1bJG+`$eh z!4%e7o-cz*nfW>P$P-#0q;d*2adCGme>yg*&?mbz5OBBV36)C(G1ZSv`+>19W>y za1-Gt7S$!Eau@Mi1{IOdm$y_0`~`_(wM8@@861HL~`Guw=MW2F~S{S9KZ^M+w}mS zFg?QwMtoso-&58RmO&c=i?x9FIG#|Qlus-Vflm!f*CAs%t&WB zvM&3AzX#sFDFke6LL7`|Fj6ovM2ZK5BNlT! zuMo5l;Q$l_Tg?lUEJOHUv9k*Q%&cPUOcaotVSl)ne|=?cbv=g49ejWj%tw^y`gh)j z)HMu4?(Yvc1|ZF{z|`bm_~fcY+TO~Yr5#=9u9nP~bqFHNAdJ<%rzIOk*zywd0pq_d zM}a_JGqo;WUa-2weR=PAam{Pbyx21A7p6rI(>1?niXeXgWqGHL6PP><;=SH;tg<*` z?_IY2QfqyUwR8T7?I8^uijcqMRAP0@h;z?)=bk>3W6`Pvao?ux@T|&D>mE6z@yw#@ zAUW<8nfgXy`t~0BXb<;PAMz0mu7NkW%n-$li$QCA_uQI36q(0j7rC1A{g#28*M7-RaBFVIo50j4slV2+yf0FyPm~<5tuajl-&70UO z*RqTGMg9(dD2?N7&Fn_Mb!KMYKt|2|6{qwDw28RkT!5EoCH7UZFn27{&$0RbXYuv3 z#_z|dK3au~$~U}_K(AdE-TqN7jqdY}i5u$m+nk9T;q@Dl3$Sd(oV@Ww!E4v#i6~KC z&di;Nixs;Rx(m?I&AnwwfviWeyp3?>v~BV*Xd&v%3!wk#-N-G$ z#QYknO3Fq4zTmh8R64v^x0u}B-U#Fr%(7tv_`MT{<%9k3e7Qf6>yvDjO+2W*mBLVG z&Ez95BW@IZ{=zPyuSeV@T(;bB6ntEgDrODK^Zip)11LM#KYrtlCha4e_e#bQZ-sN{ zg)D-vx9{&fjSj-+R|GE(yKPljz_f~r4?v=|=9oyw@)6WTWfzH;i{i%R!l)m{Mo{`C zr_jkq$SUY>Xir%^TX0yaTh=bj%(Giq?>Q2eAzJ8ihs~5Up6>Z*l?JgulrdjG;J?Ds znT4;PE;v|5i;31pp4gJK<|zDSCvl--bj$PDC4HjJ9VN%(6Pk_mM5Rm5N!pN}B-aWE z$h**QJ*G5%M%uFw>bm%as#YZkU0kq9LxFcsTw+G+_dZHP)&2MsC*0d(Ze-Os*C?6e zVa>HO=&-y2>*|uDezOJ9#6POzZD_dqE)rDi@yBQ5o!*qd-r?XY=g69z03JzZN%>gG zA|yM3gV}raK*FE(-4 z0p13jzh_J@h@&1@RnZ7j_iLdicl}Hh2cHT@%G98ff0$Sp4I^hV8A<2LGiEa6xI^Y8 zOKB-v{>XU>uu>?!qP<(4B_j?$o$?ZsdSKcBGhTbWpRf7Qp01~fb{bbFC!Y-F#6WZ! zZYr9s$u}rAFlEr7JE7Xz?KVOEPPuYo<$M0kAG5P@& zMNB&T@wVyqpalb^x?DJPTT&P91h{W|K~2j*7T`?nOq^YuObu=SGqX3cf@9)jAz~!@ z&y1IsLCn&|#ng#G%*N2gRK(QS-o%ta#?;Q-#e#^5nV%o-e=Ifsmt!hXW!||gzafFj`ivK zU%$YKe_Sj4hW!a!iboAg4O5oqWWn2~B|4;L)7?zw1QjRLUEN|}yR4>dDP6r$vSFV< zkT?zWH0OW;yuW{buV23Ru-wmFPTlj*-UD(D@qxSL>^D0zI%N=BHBukq_nJYU=x zw47bbPW-+AC_-yt@!Tyo<(s6!txNJ3$!DD=% zY0J*_9B^~?+F@VTRj_glJb|vo(smx4TZi$pcEN1|2ies1d#vWcgw=g$geptscDs!F zoB1-*E38q@6$jl;Cm_{n%iHZg65jf0UNa8fSwVgvKF!xz6X{O4f$7ucBKIBSN?bz`1<^>2MU|t-imB*2yx;Z z2b3AH1iwi#Lh&pRI^eJRQ9}hpISSlwoF34!J~{s=;>q@34HD*Kml?k>en1JuSWt8j z;wt`#wWw$P;AtCBwZQel=#C1V*Kr~2g#PFG+Y?fQ#~*#1-+HU#GhT-*E23V2JLmh4 z0|>?)V&6Zia_PWPAqqclHDydwV_=5^z+oG}CB$;e zZf=$d%|6ctMj+dUUr&F_|h+%o3rK+n0->;eUvMh_wnE-n>{ zT99%_)DDkIr2R9(G4kMR=(|7KmwTTZJf%1#d%$el@ZgPVFIZ#dMKmlDqgsOMZNjz{ z6#$#)hRIE>hYhWX7 z2kzsDQvkP%8>qHj1T84}dThSH*gOWNC^n5aTh3tTHi8@*aQ;v~`415i=d_S1y{D&r z*FpC@#v3V2Eim(%-yDT>r;OaOU;bZ#Q1#F=Q+?@lZQ#2Bt%s@Oj>kr31kSZmYTK1w zQjZ1awAN<&v^~1Kuqvgi;|1k6XF}5RUU`ooK6*mE{RmG7AUc+;IS}d@aj6FvH%LCn zU!VXq!hza-)EDIst}hP9gw1jHG3USNPFYz7A2Ny9qr78X#c-8?!na6$#=i^%bO8P; zTsg#(PyasB1o=??7I;hL^M|N_g2X4WC{~8#eoXdn?O)nL(_dx? z&UB4QuaNhI%iJwc{POkC>kMtmJ}RFp>)XP;GzABD2i?QlgUx+Wcn&$}CiHP5V6(D; zih9vr`;9o&g4O|I7Xb&zASaii?{m(lt`8VWM zqXD@k1KSjHtA6f>(kj>u;K)g!9@m6PmFXPg?l;OOkH2jQxB87(a}gLB;(>atL9gKV ztnIMG%bZ6y7IQ}S{aLpd@39jzfIvM`f(MSduw6rpc)}#zp^%F-v1N>9?L60&e%~*D z9}$lROs@|}F=PIWnC(+P70mqEfUnVY{=74fJDv|T&lDLjIRGJ_Opb8p96=Mr>fkn= z81%f~obC$@>Oh!3ki7u)wD$Dxs(a-GLJS#QR#er#SU(JYKTD`SX~cf7|I#;sy&xgu zH|S0;jQk+>1LYfO^job10ko=U3T8KA27|vF`05w-6*TroIW^eN2Quz~zwu7IWIe}NFlokw_3b%S{%6P#!(7%xceqX(?H}$q*;RksR zhZMF!wOPjN4qhANnjqH+nE8#U7Fr8hxero)fo6wlk_Kaf((wy_L+BC0FDS#_u|QoH z;2+mbN>bc|Sq&2`fpgb^Jv?dT_3L1bSME`B%cayy5+fo#XmJvT94N7F z-n$4|C(r6Ot1uc3m~kGstHq|Bt0IZ9R9hquJB2^%uK}tuR`&9Xww9om8V){V zC|$ZoaAP+v7R<6O%psPr_0K$imN?va+g{-Gg8r0gvh6?kip6%`81Wq3pK&o2u+#lT zcPYNG22hqd3)D)gy=NylN?Ys1hUdF1^~~Ct4tajMr^hz`i;v|v%a2uNaz?HVDoeWHu=9On zP<|IlB#V~?qF1cNW#(*Ddyhi+LGn~NCC5HuN$lCB7Ut2qsD7cL2Ys87HKzygQ0=BAA68v0y31oY(7IjebGs;lp&$G2w z=X{PhO9t#g5pF2F*dI0UAg?rB(VWz7tR>& zd9BK3wS;CBAs)JzfEWTj9@!Pd7Sknm0_1hjkh`xHam<3)=D|LLsk8}S2)_G0K1?Ky}gmT;(*cTaa3) z8kMKLxfeiwXSa%*)!_bjE(IcjpWv)gMXfx|uGx3lQpjcurj5RX_ZvTyod!B-lKJ%2 z#iE{r4X;IBPtUEAg^#}bZeYN^EPQc1(IY-6FDmfD@`3<+K49P6{{AVhw;Pqf2%7+M zOKNV-{Fl#Uw7H4TkP#S$7UHK5m5q+%5vsWKS;Rdc?S4)!F76U0gLU=XTFb||qdv&?Yl zusl~lwF+o)z=%AzS;J!&&+p?$=f*XP%}&vLK!+=j1(gvpj_SOD=M=C0wfpgy?G#F$Anwf+%m1MuB@CMHBKCJ zPQ8i)9dmW5WVoz*izNJ-;-gHkv;aDHpaAX_j(6XmJO6a z)skFDu<+K3uOuP=xT{;ME)d|5$vY4;yGUY_L0pN%vRBXUOJof0=z@`pw4u zi{LjZ7e0%Bbpw6TxK;Q_wZ1D2-?z=GidUyYkCSpS?vTYtV2G_PJjPy^QgbsML`R|I zU};*=B{Mh8p=5*h;ZdNDN|M7JeJ!6!T*prwy_nUCPjHRS*@2P#jah8k8Bj4A@dUhx zqlB!AKoWIm(Zka&gvKv|Hx*L2sN^+&aOQje98fqg21rRuA%KuX6+S3%qrz zi3AV5B21JsQ4y~MRg!DjetilT6&z(~?Grq~nSJB;{6=RGtJB&g2e@kF0nP21JI$NR zhcoTfad<~&G^9{ib7nZP0nwr|5zcJA%GJ|lFLZPFHN0TwPG0)jsC^hLH|i;eaTOBL zHh)c>a3fwc{z_-!8iRZo+)w!WqBsT-dgvx4EUgDjCc4CtXgC3I7|%m!;#2 z3rasu{HourRErK~{bvq;tE$Pah0yi zM$N3sDjX@e?oKm5Q=g|bw%BUiSCIkUCacH<9Y^iRw+ADe6~gpq>G`u!6qs^jp*S^E zx8v+TibWY~trKtroFuYlYY$)vj1``4UUv8B9rwiJVZTxHwCJ3Ekl@N% z4OMvZT_S+3;2E$s7_hpfsii+x_VN10~#*vR73$*cq&XF4%XhP$O)zrqq=<5a_B+ViWM z?D0B(YniZUGWsBQdRUO*MB`HZgY8iQy_fEDipu2#|BB+7VR-Q7ja9I0>55gaH%TxG z3hnp;vz4M&`h>W5E&M0CaaS@6Tw^PY&T4|0TUd{;PXq}|vkLl_i_RN`rY`t%?k;vb zEbgst7M0P^2hI5MNdUg(0jWsBRK3=rq|*KTEWbv%tk{$h(mBXRbrMN9O3Mu=-yt)P z5elG=F%d8*T3u zV6Z;mdX(I*{ZZmJk8D}&42$54bP?J`CjZbWo`XwvQt3iT`%)XPv-+7cr;hjBW2Hj+ zUCU>vWCZ+$0K3ZvtPBoqC1!rl&Osq~T$@r15X&op7-!6Qg5IHZsgRUQm>5$v>Td%+ z9^}Xy1$;P)0qF#Xmopa7E)hcr|)6;RD?b{cBs_mAiFr#yvxM< z^s=Zf+f&_gvxeA9+oN7)QI2D*BH%H-rTvW%oUgY)2*%eu=I=IMg8}|h${3^)ylSD@ zEEuO^6(m@S+jnmQYKV&vfMlpw&0>J=g_H}KTloj3K`%zLdB=Y#FrzHjnAO3Mz|dK9v2dD-sMUbdgq-f&t4kX*tkLgn=PX91-U_>mTtT=mP5_V?-v=2tP*^uuaRF zsF~F-Rfm)zec<_2OzTxrhE>Y*X{w*sA10{EXZq>A67hbViHdqDCg`aYVWOL!l22gh zmfKf2<7%gX-H|p9OF|F9;zfFO-kjIy;h7C|MQ+q&N#X2Er{QOj%|r)_5==cIN{8}K z?$4OJZ)?)-d|Q~`r>S%q_n%G}d}ynqKcmTeU#w9LAFvMwarC-;o^Y_Mvloqoj*vlq zHaETr4E1p<8+_=gJow=i=@=W|D2?Jk&m$pxOSfp-y#7M1?=>^mTBw7Ai^3#eguDceovs-$04%EWAzLqILI%kY@83MYo`%|0{i9q ztWWn@hNe@HLI#$CUAbtz5nkF@j#LF)bj9lSV<;zq$o=`3-5pky# zG{T?8EecZ|)kt$xPmPDzj#MnMA3mP^)mXQrx)IB_%vAJvHBEEw;1+DkABe4;ck{ip zF7zj?NB0-NM}Xd&wWFC8`n!Hr`Sm)=? z!-9&LbZ{56%KI^@-l~6l(}sww<2HFanArhTTh-w>D{0w)yh82O_Srs3ZjtM9F&><# z!7Qf>(K7Hsrb+7o!%yng8UfFTKIWf3%mAIf@LJ|%yGYh(X2BYrrMjybc?QOfqL5Wp zanz6P`QWxgayTkuw`%Z7o(XSnrYRwIqMnF+Hj$i}GOGN1geU7BL0!}sP!yiRd-U=l z=ndO1sE=3F3I_}HCXThtBRda*uO9^y6~Psd-$=SCK8WqiVk};DHf>Of)VgFZ7=V+M z*4Us|$hY@=nBDA}#xw_huUF*OYQ0b50@QQ@P&bNb4r|J8heu1uM|^wM+yTz!W1pS5 zHlG%Qbi)UvK#W@aU#$;Q{14kvsIe-Cg~PoHh7=YIYZ~V1nEfM81sSAcd_FR6+Et!* zBnOYEblj*cipri#u~&Ah@1y8OeP}Y%^26sKm3zFElU_ZLZV5(ibQ+8`HW5q=^pu)L z@S$qC_DKQhGf#srf)H@#N1dXA=9Ze&RVXV`o59r!Nuh&aezk_Z0ej&h@syH=)ru#p z+=KMJ%FvgxOMsX`n`a*X5(yU_vh?8i-~pkuV6vZbnUI<5v=Sn4NpOYZW~>=wIHc@W ztC@~iI_1=1XzaKF19IcP#E4W77+;c!;K`x#k}$HW$f4bMqwIx#)+83kuVlwO4zUkE z4ap73jr2*P!~a9(Kwqw-T*TdO)`EaN!s(W7=adTv>H2SFK)f`udX{A~LRcJSMV^8^ zREB5bG~q1KZ=8B@iS|3@?+FbMVZ^3P)ZBA;ZiDiIz9|ZZ{Flqy)52Z)_azX^6>rL3%4JZm*2}|kCR333gpIF zrrW5-d0al+m4;ZiSR#g?RhwS(+jEb!oGYfnq9L7v*PAO|_Gs~0^_8>HJkS+hF(Gv3 zQ+>7BTc2osL0x_hh3LqHzypeRsCwXJM5!4Nbl<%@-v;jI@egj zj2Ma`?iE9mCGM3?ZkPu_6;#c)as4R@j&hS^VbfD#p<*1RvMz|0sA|I1-wExhr(Btf^}0q6CyTe@YCT{Z{lfEE6rVfN=huww z6)L+Q?&_+mjdiZG09Vc*hIZ@K16RfJeW{X@i9oFESBcdDU5wZBcIR5F0byjc(v_uU z`89UuubGlWL~i4k=gZ;v${+n-q+?87mc3$`&xpxZH9EL?e|YlqpDnsvSITp$V>fv> zyIbFuXr^iCE~YMMTQdj`2{?EUe5PMZ)F-(Vs+O5XTF2{$+m}1n-8CL{erHC|!PgwE zV6AwxYUQ%sK8iJx9f&(e5C%z_CFNrS25gubXg;DHLv({q%%|2a&m#8I$h?kFbPCht z_w?t}ngr?Q4f9tB$?NEmxv*P<2k-? z1`&j+RqwL}-;W?5*`hh}Q@YR_qgnFWprwYgNJ68!Ka%vdMoJISD&poA4p`(LD%qDx z@fHxVS%uE)UJl<*{Bvk|nLPjF$ska7rf#kFeV?F@`Ah&nb?<5W?DxD@2n~x){M$FV ztJioxmVZ#|_PFIWy^$g?a0sn8bRWwm&dgOb|GP9jlsVG-}9do&Hm|9M!HL}6eqXVgc0sPZ9iHcW9O%h8g?8oMHtG4(y_ zj>OLa+!TKo+vQcS^Ky=xZhz7?{?j5}Nc`?}++q6a;c^*$(06;~_kHaXr^oiO3^W^V zDp`IO#!K)u@HPEm0&!=i5;x$Xa!|*cqHfug5?z*AWX$zv5v)Efjk+Y}%;@cG%Hzsq zTj>htO0C+_*GPO>a9QxGvO>L?u#$fz$8gr7Sxgl2O&HKgIbx=wh7=FyDY`|-BDm3q z(ETMFqmL-YED-eXsv|%M^GaG!bo`1nFV7Kw#_7fyS(x)hJ@&UBDv8PB*N9m!tdl2< z6%xe9bkl6#~X)I+$xLCMuc8wQ-P?jN#`5lq$KidY-Lm1i8%rY|@ z92NGu10M{K-(b?LY*a>&n}qO$G}%f-s(zvask*ymR16B1XT!%?2a}3JqTI8xYJOj1 z^MxR867>A;k1pN+1zkX*za&|oSQ&T1U=1~<8zUa0jt-pa+Jy0{E1JHqDh|C>NwOcDRmhsDUYjtZQN0=Y02DrnK!?D9j?Ko+CLQZZ=Y)!JeVmp(M zaX;W6*S!*bGr>2ir|ZHh6}iW+1V%*5P`1D;xXzD6(PYb_P-eC_jiCXCgU_#{$hIy%FdG(qr+D>Ix% zQF*B|O;hPKH>d#CCP4BSzDO9+e$-i6a#$iNsBKhGYeGrRqyzFt;}H;2KB|{AQDh7Y zV{s^G61XI6Y2$_{GKCCdkvgGhL=Z;`bIw_x^b{qC|WMa`M8;n>d6+a%;xcsS=eNW*U{HUOwoG8#a z4KNI619@%5EwZIm1gr>X!mk!Z7xibjD4mVz^@OBH6ede6i3m+ zR_HfMhCZfI!BKlhF=IpiQPAg*$kQwE=dQpD{2rdcAwN&AAelpS*b^gHNGMlG1QUA` zj1=-GppSg1lneO_B1HWwS`ii978)W-CK4`}lZ+}9)M{QWp|QePGPb$Kl*8~pKZkSI zqqgogBDZz4?P=?8Q?-RkzU*c*5-|LC2w1p&+~_rBeZIP^(7SrijBKD0ph`?o{zmgMQJ)Li4GmXlhX)ymo)?J=!tJ`_aj4u$DKuh*YH z8OA-~V_}jF%i;OqP2qmn;1s$8p;HR&QzP_lq4GlYB1~)lcLGv!k-z;n@&MVQ|Ahj| zGb`8)<|VEJp0yD#ID$rl-$1O$kEVC}fB-IRtTpH|y2gokqTx!z8VI1=kMB3^89jhs z&<_|6H};P{i!6I!E~s)=HR5K&%*?FD75K6$H8M6ff|Iqesw$HIQp6w@Jp zY@S%(*x;|{Fze?)sO0N0gR9U0Id!(DjRr8y z45-mS9UMUBMtY+Sw2=XnVt_os(Zac=qrNT{B@R2!sX3248jnt!grGo#=HMQ1tGk;< zKTl>UGnu=j9ZLuXCaaHu?oCbfJ7R{VQtE9}t5RJ?u)#T+^Y6k=ix@I(x-qXBp#inhyj3n6WCgVYmO?T~gv9*rGbI#Ko!*I*L1q8i8}5n{ZnOAmid1P0T5RK4y+;f)s*vS^G>583ogms-Fx|tzYIy_xo9iZL@T>xtW8v-QT&{o&c@EGL4u~A(UH-@YpZWje*Y`A>Y#>2P&>FM}5h2Q&EGDbT2Ig>7Q^%dJOsYotV`7<( z!BJEQG9iDh2sq;wxfS_+A>jQApV#K|c~M;*LTP`%=Jy8>uJiktAY{nZ)k5MzEb8}J z%#hMv-{2KI*wd^%uKP$wbe;wU5BLIk<%Sw4O3~OI`G!ESB2z_=ndz~rlU3wc)d{eC z-SrIvcovw(Yg_}katjrUE=3uET%=Z~D2#(4QmkN?A|)j`DOGj7-}jaw2^X^ zOj?TTx>i#|2=#&#{NMFFq(7Q%&azZE>dqf$<^=8`6_rg9_t$Ekwuk=Z8v24YT-;U6 z8q?D%bOa8jND}{?xdFyC>{@d6rt@Zj!LL%Px6ZaLcGk2+qd1t)=q_RAUY3hTsm;^@ zcI+2m#~M%oKPBgu3CpbaY2UNF=YHSwzVAK%pDh~R#Rr_krFVLqzPJ###%&cIZGak6 zPFmQ@aL{y~(MxMJCH!KF7ilj{3(HP?FWJTIQs1lJX?T=8s(+1rP4lY%J^Y@*KvcY% z({Ngd0TQR)X>j;8OWjL-*RpH%Yus!6J57gNhy3q(zfkKgG8!QT*-^`DEIN1a$~ns1 z4KPo3Bd-8bHdSUY|tz!74y1TDk@$dBUjUWx!^(*+$YQwK|j21l|5DBznx0bdgi z2vL8`7Slx8nA_v>K=y#a5(PSWqu8#7mXm{)S#OMDgO5Ooty)JEc~nrOQcdtV=fVV| z7(#Eo)W2~+%~=ZU!0>5VXDJYurJ#o|85sU#pSb`K`>#;3bjwiCz}7Q{0vd6Wbe?mU zfM$hd5oUoW;ZQtgh6oMtmRTqna+X>FY%3?&VfXBOy?A%=+pm8QvZfoJT`+z9>|K|& z%~`m3H#=Wnyt4R|y_6hACJ^61qyGIsv~_UW#q{E-J6?h()l z9uVZ?M__TszxL@1AsRiJrz=d4;hnBC&NHEZi>jpBhp-+HohSmX<3!0og6Q$J%yYB>U5$bQZ|g~#>ixJo;N9^cjx%ddfk5=!H1<(Ao=)t>w8j9PL@lRXoB6T}6+MIvv3@GmE@0;kGmF7vMSt`%>Q z-txXFoye$nIezK+x9cnSR~|N_)(?HS1du3_uCl@~g?FGr!S<``hw48wKwAwqJl$H0{+Wjm(wWtvTGPK=Cb2ZA|V*n=c&yafdhDyRbP}tnG9JnJ8I7b`AX~ zw`Jw-@6h=bRyC=@plb^_Ef-wbixX?Z3}?g|6QAO;E|W8rHl>7wIU|Nrk;=McUFx#L zt%+NcPvi%Z&*rU#b5`+0*^cI#>VkEo?ulxc8Rv%l!B7xG+~!(&d|)o}2oRTj!k$Q( zAg@wqGWm2qlWL7=O=7p{5#2%E%O)(>2QO zMA}xG<=A0}(J>|xO~g_wbjwZab?Z!96I)XEn4Z=>r~9eyrzQh95S4gZFCZXm$%m-8 z^OcX?rc#ossAEww=SFflV(}_+UUo-4^P(={|KsMRYwdnH^W@)VeyjM;x8&6qWrLnk zmT0u{tGjx(T#eCBb~9i6E!T#?(+#$=5WLDm$Tu z0GcS!yzEW--`*w#sz_#59|=tx7n{Gt**YpcG^%7>Exfh)VrLAJ4({5xDs1`cv#)`M z$}wujy?9lLCfVVW?888lD$K|axlC@6xU_LGDQ%GMFy1Ncm3}4-OBzHo6d16)cj!*fOc#qgIBdDM8wNOEePnxi5ZP%cqShN{4E z8!5RETuvv3YwA}-G=jMzv@wY12Zw_sczrf2XFIYz**#e{t2XgLp770yL`M=Qug{E_ z|D~KYWsd1B<&h4Slvip*DS`6={OGY*)EL!8)fp78Fh~N-FkwwxpFt)O3QFXrROuY5 zn0i-(dvCRs!auGI{>lm{9xEaF5-6lj@i+vFsb=D{`y_G6G?OzLt*Q#m92)xkm-{mE>rasT zt|>s)4bWHsjJFTQB1~fy67c0HPb7ar!a3DWI}s$@;KT*za_1Ay0jJ901c$TRO&3x6 zksFk|-RL*y)jGdE><;=83EXP$i-OzH|Vflqh&$?k6 zwyRl|4@2S;%>5d`_s0&WC*~!m(*sKstB@1lBwIpy8TNJSvED68mxm!vcn*Z8JnpGO zL#}CUOU8YiQdqfUozjRrWIcd!N@om~0&k-dA@p1j9K0YzxDAX_`Yman!>*9poCmB4 zc_dWHz#suqPhmSpCj$w`6bn>|DT9TDu`(yn@nZK)8kL+`Ooxk^S%vAO&oL+D6sllh zq(18i1yd7XSoEzmN4`3(s(Wdi27-hpzomNRP#5XEVv;9Nr8h_Ecvyyi&r@GZHR0s_ z+cp16kcqD4?k8LWuGih4x<5_u1sCQkonVXBq3JdAYdUf(k=j%vzF&5&Z)V*RWSpu|;#Udk=cd-Ve(YH7<7o7&MbBpoyXJ)vIB>sRjw) zX5ECa#_i@ckjVuZuO77y;w1$ANd@Cwa;iElesicI><@v1OwgPBLeK-+sXCeUSIeq| z3h4%hH}7}pw15IP$jcKTjh<4nnmV1Ruq6@|dZZJ;64Rw!P8+!AG!QJ-R=C{Ic5^&m zkw`+Pq!U)^sDuiSn=aRJakD`b6R@<0uIAB$9iXcY1p(L+ofgt-a+GDpXz>f6ViTkZ zmNB_V3$%sdV~4hyp=)< zXesh=NB-=SbD5ye0p+;^Rm%pv{Zhm(y(@dwj1(RzJ74QxG zqv@xc4yRvDzaM@t{eJDoX-zHP#!s?Na!#zB<6Nq~2kowX1Qy30QtNX(-qgHHb$|MS zY87hkXkO&#ZeH!Y*S;4&GU_Nk*{oGNI-0L;WZKolZm|=^@Aa~?@OdMy$w8!>PgPc= zDx;~&L{sjm+;cfbl^dO#n!7%CNAAJg4|6Z%exCbnZZN0o0WWZ)O&wM*QD3805p|<_ zih8~JR`rAGN7b*Zf34Q&)LwOunz32bjLQ%Urr?T+rJ2TdQnM3vWirGiCsKKnE9jc< zTIG7swbymj#ed}btLrSx#V*+>9tSedD8sZ>>tOSGDz!6=D-hEPV+qS>f9 zs!@fYB1j{^Y(4RQ!|m9yPHY#!`y`|)_hD^KO)4i#mbit$#Y1gYM{60gm7$n`LGH@K`uhTR@ zHFWxvP(o*`Ewm`7heCp}T4(}?9nGj)U1HM%dY9it5KU*IagT;kK1nMuDiavP;ixWF zh#3Rs0Mdsv0q_ZrV(J4(@EHPFD?-trY7CUKcS>_$Ik=C3%N0)Vh2RJ5iG8!THnz-O{nnLRb~}w) zo5AA^)T|uWF;}~;F&=iOYx;IBpT6>`J1<>PpYU5;_F$@_x@}6X{pRt#W0O0J_sC%( z>YCI#`5s)DFr%)%O7c>^PYs_^c|k`*-XxwOn=Iqi$SF7pcDc=wAY_sMP4-H$%?uxi z>2$``CX=9pd|82}r1BmMjb!YbT&n~$8em02N8&_+WD|0tBhi!Clju(zOYjM!5t-aU z$gWD7EwX^KkXzFs91~7}4d||z)~i@krQ}plbNk)l7J+(zT_L63=M2&$z)lME3K;<$ zRw<=xqSQ4}?wUBS>zT6MES#c1J;gxg8mOjnhE`N%V55;3&Cep7CI&84l^DlYLPCG-RV{qjqC( zWAt8jXM|JSAf4D;&S5VNEnrRX`hs37d%|i)GUf2zHMe&^*}eX)o2RTP>=t=#Du*|7 z+9{2>iPd%Su>gaHhSv9hqH5(8I;lisW_LWu7om{%%{kxuOYiOLR;hfH3R?N?uSUq;fUd|LuIklt0B)XK%Bv}M(z53 zFz||0!rvf62k`f~AT%yHtWIs#1Ih#wcE0R(t_#O(JX{Kua7B_v9EmHbkkAM;GU4f% z^n;P)$&@6GR{KHFGD<8NDAtiVks&>qJsFY-f|p6AG9%mRsUwnhLQo0rs)id~=M*WW zD#%glOlfPdWIut)O7k`ZWfevuPKB5>SR+v>3MtI7QpSccDq3T4lr%sQHHYIk zZc4?=t|yR`q#`Mq9z&0{N9;-VXO3mK9^*#K8s|o-C$S;D)!CQ6&#=?6yYf*7q>WV? zH<@lV6FQDvor;8KN)nzaOL(R%;aFMl)KwB7faA3#gJMLPD)J@OT1$(L%7b3d{DM=b z8;aLVST%nCvRTUxE^A%ZsL^M~Zkx0+>WXIaX=lZpY3!7*-nzmT4ynRZzjIOZo||6S z`T6?%80=Z;@cWWOTko_5A9(1QCu7#W5k6qLK+fzaglpwFoMp1D%eKn4%)Z36&c;Wz zkHPx~-Y~yS-e%r6yl?*p^Dl#TqrH^Ke-X2kStYJzHi|bhTaBL?KCx?(>R|^~Ycwg! z2q8%P?_xs^gvUDuaK!;{%*wL^IIv%@(>Un(>R@~wvRlkMmLn=fbhttEmCRsc9w~uJ zb1m{@#1?VB__?SOLy3|jt41hGpHc@br8<_)D}18|hB^UQ*c~oQp%U^frO9Yg_$YXOrEY9o=V5c35vQv^eh*NZ!VOc;) zB1`ig9YPC?AfF!*SUb}-G_0B1wxDTYgE(bi-HDYK4Ly11+kcayb}1ij#9uyp`OMaf z9lLMbbK}v^u>I5TKXh%-lIz?pDSB`$Bz}(tA538V&=sD^p?pB&7^k3+CJtT)^Go&i>)9|x+G3KbdxObWoP6p)& zb6xMLUVubO>tjgrKK&CIKWYwijaZam=!5>G%9O?}X%4B9*jOs5m@6?jDpvL2A%SH@ zW~nw;8_$2ON~vBQN$hsob9VfZEp*Y)uUl$ux7>o?Ij~{vq|y1&U=j;Xf1LC|jB@Rz zE&${a?9EOgw=QhU>^L^BAw9M(tnrymcC9H}yLatEnx8+lIG*_cWGRbA!&`^ml4nE( zovEcVx=nLS`tHO*)e+6U#G$G$B43WvYIB-euD~^hrm<=e-Gn9)YzVdoZ&zzf-UGX9y!_s`JbMwT1P%}Js z6v#)lu|JpxU0hx`aBcwsT{2K8hs7a35!?%kA5)QWWko=)zQ$P!KG4)3W$nOu3kdj9 zpBjbJn9{^u-ind$;>P{eR@bVdlepFC(e=$g{KfR%WwUR5{Du>A+b+Fn#Z_CcJ=xzi zX;jDbx~7hF=$fVB!kQo4_Mpjo1@pj_)fIJ(i|?MvHYOqv43)RczAao`eQ~yGqFY|w zc2l-`&+=Q}Xuf9Po>f;qxPMIctG}CrwYiy-THWS=gX)v<5Fct#a{N9zg1F%?_UQ^r zkTo;8HqVYHq=Uv-;eBD8gNt!0aO;YwGH4J)p^~%gH6Aq*FGki#&^SOokj-K|5`+{V zO(bZLBwr*L9w5Jw7fBV7V5KBsFPw_N&B%-4Fr-9i)j>#dPui?jeQZvnfS5z30R{~d)JT3 zD>cZDQo~ItT{Zr2KJHFWE5Q-US-_!6%u6U+QJ~Vmz@vKY++Wjq`AGm$001hx8LGN) zDIOB060B=c@^Cb;GR;Vrrzu-jSVm-|TH8q1s~Txao{@5I*w*rOz_E*h$ko8J5Jmya zUIOhj>wqS7j0CE|gJ~W~v@h)gApg?TO>J|oPgIO9#%kOaOUhd@rP5?=EXEq$=6EyU z_CL;OUA%2i@$Qwid?dn$JxlOIS2u?1+lsoyZc!bHa3RM^=Fp0~I;!A`Boq?lDe7R= z?+3CYu*vLfF`41jU_NjYSp*;Fj07zdQi_I1&`cYWD-!fQhyMnUpQBO#d|iGohdFpF zf!@bi%vuelf)DnVta6!2==CN;&_E1Hms5tzD{FH{)hbJMLsDs(xz(xEaweV4_dD@+ za3NB&(78blbOcB+&>h$l=ntqeftJ8_XdDZi3~>Hw$3XSK(7_tb-g5&7W_fuIY8fo2 zko~VkBNtK{s2UqH_Yzs2JNK=s)*{~=uvLv^FIRTR@+HN_A@8DkRU|@0=OQ9PGm6OY zYw#w%A(-!V;KBJVH6_200}fXn@;#SkmsVe$y}G*3c2oAI>i%qh^|9Q^oGzyz9-}#r zs)cG&83|TXW7jCTJ`;CZOx1dh5|O%cfpV0X_5p^GZ8Yu}@@BoVx5L}*?eT8%YG5T6 zzdx2rDRg%JH#&b-nNspGz=s0tbu@J{1$MWPBB|%eyQqfzNtw;+vG?x%1!1FycBWulz9AE96`(cg}5TncMl6*?rIY4cClL#40fn1h+HH zY2hURsWx_HXA7lnOJ}j^Y{M;ME?vENLgkorS})kulG$7lwT)io93*4&Eg3$l;44@- zq}r;)J^m`Ul90V{guGyUpL|OGV_^Lnk1pmDMNz8v%{DAHTxD2m-ekDLdzax})4jqI zo_&S`ruT(U1RH_%^%{@IQek1s;j%CeBonqoHiH9xlPaoC2dj_{6d-4FiqT-i0V{lH zfc2GEUTGP6<=lF^LLoEMh$TZp^t^rivPE_OYs1^rwXS}yM|6t!BY%nn!xjo95-a93tZ06 zZ#%E8At4C41%klo0Lm+R5->fh=t+S10+n%x1ZQ%r7L9;& z9o2%mk`|=Z{&YUCD{Dhq%Mz%=ejVNpFI`YP{u_dpJ(hk8ld-f|wp)%_PFh$C?O&bG z)A~@lDsNVlA)U@H(1s&vpQRb6S4qvdkYfA4R(VstT60d(F|RMAicXO|c*7`n4dCVR zWSShDK(NIUl(m6+waJQ_(D zy9^jq0l%{YVji{cY42}8)~;#?o;R9IK@%}0J#K(0Zi-1BbRTsecQdkkJ9q&<4U5UD zG#pARhtfyW$J30Q-kv7uPm!rU2oB-o*f9!V_lx=NF(=26J!AUE92>)Apg1vx8RMSN zK0s#f4^t44n)Y={tpa)>3|y`u-&2tXigL@@Y%O%2frrPV* z8CWhG^Tn#!ssQGAzs?iDdP9b*fwU-npk&UD1VEKlT@z=mlP$rJS`z}dAsf_$<0uqX z^H>S2qLSk<(%pWtopAbyKChSC-_^0x+3D(On(4Y@?Q8>?&P~_@2%9SVp;op{y3)3(d|W7>2Q}S(M)!5bm13Gx zuZq3oMfv{*K-4}gd$jrg#lmk)y?NT)_2G`YI~H7(j)SBZycS!^pPG|4JI54#uujjG z@m7R0wQww;XaVN24YOKjT{O3&^VWNdH(s6xT*St`3-GtE9~W+EDQcI1;}+yVs{R)K z_C`5k4^A#>7qxJTVz`_Liee~X^m>qs6j4!({>edILBrv+;+>vc+mXgB;PWWQ{F=PW z{K~`FxmrN(%)9s_FR_@6pkGr#qYxHSdre1S!Hw5viv&$2ydDFLE{R&e?+RWIQM_)K z0L7Am%7@H z+lYk?n=~2KeqMm(LPq}IJo{4n<|&t4YZHJakGdD(Jy$h_HxzZt8~6*@u>6jx!@!3I$V8YWyI%9) zxE@D)B;c+-J+Z9^~BoTb4l`Gd>$ z7Jd5^eV3LpKux80N{z!1I3m$tAQT{+&1#1=dtA)x^Z498hGvS!X}(1OJ2aL6a`FB+ z%@mE}0Amb5o~Sc`e5^Bm-dm$!(H>|_P=?hNxBxF&o+zx-vpw8K{YIh3y@}hd-!5!& zzfN8aYB%!WJvME0ZRa-`Hkr1&)HJ=Vx05D?mJ`|(XTQZMDnaA&>!s?HloMQB|BEY@ ztbgyFQ=cBsO>`P{?N#YO++d4EJ}@=BybB?5hZM>iCH~w zYys;F-T)BPaq#wp1Ww3ac;{1TgsaPb{A$H}2y)m_eWIFVtL5sB>Yi#NS@ zW%aVYLw^h$fvn!0y&z(yFFDEeWhc4*8xgazqhIM-UFuq0?pl3*#H^fZI#{yOs7PMm z{SU<|T>*DA74yTYuu4}Vj$;8RlAd%NSNNjmy#A?je5^4dw@k=OwAkp{7}yxwSgE?& zw$a_=zYZ1~Y)oyj-7ej0+v(aB*d^|cJZgJFd@^##_I$)L&W@D`0L^8Lj^r_(lhLrf z^bV8}Z&EDo7+K>E5{>sdv*U;Uq97-HOLcDIMVCD>=aL_;nA%!Xf6>A^DPM@mOUBGE zKHQ#nK~_N6+0A@FZO0Amq0CKx+I;6P|8sMdA}EYU}fY7^`?EnT4mE8DD5^0N-f82j=Ye<_v&_}h|Uc!kTUrp z?2`!kg@BOYupP3IkO(m&1+Y3_4f++pDn)_`%1?etTVs*~O`!FB^)05$=tP4OT)?k2 zbs>1&)O>i)(EP(Y_!Im|o&k=07T(yBFgb%@uO}t3#BriBBjlBQz5U)$DQ^$f}KQmH;bQ7O?K~RLOpCqHB67d+-W-C{E)PE~DEN%|I=y+>y{*c`1P1 zSqj!Di(^J2pl@{DKBl2{OjYePUTg4s67~@0^_hkuKRTt>#r zTLLkQt}Mwiyete~VgZlT$pac2H=8!#y*M^9yY!+8C9I<}@JzcIfz&*d%Z5WNU3$?Q_rDtM3crIzSPRE$K0V5RR!Y_-CjqiUaOeTd^g zLMv2MGaOWhU@BEB&I>q2E!)B^>L$Acg(C`^@NDJTl};09R@0E-$f@B_$;)<9*1$Ew zm%{nl!bc&SfHdAF1s%M~slGQ9c+%YqjQqZ0~)qx=dg@PN=C z3IqbCmcaih?%KnfI=4HT^jqQ*a8S@CmX#g8r!UAI% zOXiWJ0rG0|r3rjVHw|=~&?K7>x()JGOH2rN-!Y=av?~L=S|=E_ zPH<_n##!jL*R|epJ6-0^&gKsq(`)_k8h}9Ydqo{Xcoi;M^YjXchFg6krC>A_(~s#D zn26x&zk>!4(tmNU&>T=DIu%0&yi_gB>a!-Gq2 zaHMDC_ z{G*!7@Y4MloCTTont2wO%p`w^JBt3%WMv82ck`T1GF6&_Cc)Hi2wP3mnVB_Mq%|Ei z7dSYLhS;ZS%ft&{j{%a8P9f2*vPO43;j4not+GwIb$oa*Ht^0A8orm-FVxb zgY}xrR}Obx{^4?eQD@F@=c3|Axer}!ukNI6g`XhJOyODhczgI5aYMRi@TKvb9*`p+*WVbrir#P68DaNpy z6WvKS?|1jPc{dimxNMJ&K4?-X*s9L-t?EqQsuaeaLz$THU$Sz3>#%jiDp>PfEfdBB z>@C4o@29TVe`@_6uaUPW_ZMSk$>#mk;GHgMvX8--ri`zI=#!^;l7eN!{II}n@8v8m)&%o=qs%^cHH$O7pD{P%X_-2otXv8xcAhx zb$oxpy!-lpKYHO*T9BwPM#bS!c6?vKCt5 z^ACGr@rdfCZWSsTICE{Y*lu)CO_)b*qvMw?cj?rn-EgdUco)*Rr=#S zY7KumAMSng-d0Sc=w5Gi1r<};K{(O+bI;~d}J)u!c2lr6A|SsaW$Cl?ku?hjY~ZN8LS!M)>8&nrK0 z`|OJOT%dg3Jb3@&e|ER407l2@InL>kE8pS#cg=zKA8dx5bI;{s+Ry1@=d?2Vxt#wJ zd@!bS<#Zd@vv2~4Z_OH0jw#np?1EjB=Fjry_;YKsYIACHmt`%>S(dv-yC!3`bBi`& z>@sy{Y|h!7+vSXi+w8Yxd?n|O+_-b6w4-#d?=kOb-S5b6tmn)BLf$vNZ~DOcae0Z( zs59xT8oM@K;|%&*eeFI2mYy>*(laxO(T>GuOrga*4eVsRWn!uFjX;An=-f;Rl4a(I zxv)kPOJI1uEtM|t_{|#+MnzICdCSY2L@7%YrA(YZq!W3vPIThZsSym=;I~fkZ-plf z;hWMkKxf!-D1gRL?##^Kh%VhEaet6L#vVX#sp!Q~0`vd@4QDJbEy>Hp8g{wB6O)gQ z`Eun$;M{r22R~_=^1&}Jzo5z&q=qGE6MqEGMX<NZ>2rZy>T0^uARrQh{o4>OMcf+@oM9DUaOTmUydba9Z( z_~=yzZ|M|!=BtOYaBhrzT=n6+()?N6=}NbB|2;-0_II6~=PbHsM?uM4F55T9d*!1- z{`t!*_-DT5m#ovLW6ST&S03TE*D#jInB6iT~bi$)?lR1zT-wzjYh;&3Y(ri*`;M}uhs8w zD#*_&D9DE^nLNz1EbFLONQ)JQHPieB`8c71q-B@-iJOvGc`I}}gv?G{dZ@TI}5*}jpKCkk4v#NuIsFn+V#oHRphS?aQ zLj+DIKew0pDW1`DRbNAMxp%L;(NO2(-kDhzIJjlO%o=XKue#yN`{9|5-77cVP!(9j zaeBQi&rvp~g#W>}m!RkQ;=Gc+D-UyyZ!VZ#j;ES-(ctCgE1xV}wSIB+(%@p4;TCxw zRr(&=vG~1|TyIF<;^7aEy#~4gv{!c+Z9_b3dz_Rq zH>3}eX2`i5`KTPVNXvu_=^C1t&Z?R7SO4uG-@qwJFL8(X4F0IF4rl)yg01{f-7{%T zY4=V2r@z&(%IGxy#`qi4kojfHOV)rjU|VQ=**;_+N`F0LQ^uyu-mF`)&Sf9Y>Cb)r zv;XoN3d#%Kbc_}zi|%utmdw(jY3;?2W1@R_`h}8_(l<)qD0|w|;Ca#eV)^qvx6kb_ zW+P}Vc^99b_Tm#9_2}p)IQdV%gnt#=eK8csO5r^7#Y+D+5u(x@RzWtYG>&y4UsGuT zr>X5zX$`!3KB>}L@PDGxI@sbAsB{`mFlth1J?zM=Ds3S5@mVTu;ScZ!CdT?~ZIw!M z`0Wx^X&&eH?NDif%+cPV(i)Pby;r5R;FnceM{=~UsB{|Hs6D09dXlGIpwb4iSXZOc z7A~aQgOWLofV!+{|E1Cz;!Asz(NtSTy-}q##G}t*G}Wq3U!&3*Qlf8QbQ;y0ev?W= zSNeWN>k)6(A5duxDc3*4XoISMg;w;h(2D*QTG78kEBaSxMgIz|=wG1~{VTMhe}z`` zuh5GADQ%=}*T13CXuE!t(I&)a7_2I-ApwKLXfu_~utKFFkKque<}NyQY2S6x@e zC`;)8B$)oXz(dpeae|`69 zV?!AwofBj5+O3})msiWRV2TcCOO&=GSbgK8o0hM%wF@!ZS<7SWj=6SA6-{?srPZ)H z%0E5Usa6x<8(=x9wA&^|u2K%=>;cDL?W5qK-FV^3Y zh{k$-Qe97vv?98(I}w*wgyZ4Tb^+QIT7E6drsvVs-7>eD4g<2Xw&(cjS(-qO*( zS&BtIS9&ShDe?vhp}Iuaz86k>6bcV1HFmp%YxF-X}@pmB)H}c1mT@8B%!52AXTKtln|%Yl_6!yJ(yH z!*QsjlWK5E^-Py)Sqqp1Wl_|U2ydb4?vJ9@uGse8o><4!N+EQIA}W+Uq4n56qHiD} zb%nQ4<$|X>+|&1|7O{8FMs}~#XYlpjN>_6y$N*s3z@p9s-eq&#_=E)qZ8FjsU3fB@`^`P#^7s_)rcPmn;wg>Qj zQQjce@1bkunc8cyzp0O7IH;&bvp}<0Ge*bl`Y-V?8A9_6Y1#(UBa_Qt(f(t3v$uq&Xl+AylU!X@OI0sirLmG3+YzXXXZm zG^SJ;{0`UIH5p_#fG0MDwSe^ifrT6iYBJ;4W5b$^ zRBm2?MyHJCz;25526~+?-!dy$XV9c$U{bzxOg$NvTW7K45G^Gg!=1&Z8XA>c^-6S& z%3M85taIa90t5klfFpoQ03D=HCw{;_z!*Tot|;Fwz+S*Xz$oQqrRitc>KynLyO;f{ zwILJ%etx|k!*bb-085Yy3Iiwb zv0y)zs6<3?a*l(g5V^iyCFaaih?J)yaJJ5f<*JJSUIUvGOuAUERGBYu>4&)Hgex$r z=4h<&Ny?su+Jws~TV^0wXBR#stpJ`R@k|{C@FXU@Pj&%#upf~tX3~m;BPpXbU;At;aTAh#u68{;^&!cFE zL#c0pdW$Zl2p2$IKsQ}L*IYo?T!0i8FldSJHd5XOk6$=L`h;JSeSjxG<2*58rwk1q zV}!dba7_4$a0fJDKMvh-h`lpquu{qHkTWtF+a2@pSDTzH5z zWnU`KK{NbRHkhHh+m*gR#@lEgoDg;kcQH-uVH!xjj7yAv;ck|BEM-m)>_YFa!etEM zUcd<8A^>Iz2>pQd0344bw5F`KfbF<&3(I_?Y@H>Z5SF0nOPIEn$l1kA<{MN(W0T=^ z<7G-3?QIy=a5SPJZglQq!zFj@I$~T zK+9s=0X*U>QA0bzeE>LE75vl%7s4Q*9k3g45HJk51V|$%g$mR|m!&EZ?gk_QX9c=A z6%)!KM>$=(6C`pOOY%_p^ zc0-?G_rym>yFoJW)P^Cgnr-yT&{fa+Md*C!qmYm>cb{%w8h^6Rj91lJz(s&SPU6t_ zvw(~62>4auWX;){i#5W@(Am(%kZ|(M*)tc<2q!DfR$Rnpg^;5&w43@&yG;VMn`t!(c9Ue{_n8iwlBQwPh)J8&4eLg9qq<8v zZL6+b*QeX9+owCIgQp08T5VcTr@2(O7$@a&0Sx;g!CyxY8YT^LDr7 z=5Z9vB{w(X9(D6c_pqBEuB+xxGbv9)%F|5B)5v+6)qNV}f&#c8?J3503Vf#+-zo4> z>RNH_2>TcnMA*tG7s5(P@uy_hTw7f(|1Ffb9^n&!vjBnk5!M1?>`J5@{I?JWd7S!H zju$?EKrX>E5~!5jL$D~pwkF`SZQHi(vu)e9ZQHhO+qP}nw)^y??;E^+(H&8P98|2x zh^k4h%=|w!ei;xH#-s^4@T%HI{qhfpKJ^wO=~^K^nr=+2lh;X@B-h?9G5+PonU(2b z&7_EiTx@_oe&7Kz^kY%A$ugp-QC9nGA@+%~zd>MR57nkgVRqcLd}y=p-+5MF)?wW= zW(VT1E~H3rAv1obs2mvtu=l&ElUIBcAmhl(#t(o_w)$8?G+U36Vg zDRA1TY7D%W4NDmBnRV$i1KI}rx9V*aX3m;Er!8M&*6tJ5?hDo&Gu7Lb6yz;m z9cvG~!zpMHG@+3+vB;V>G%eJ0Ez~q6!u{JsD!-9tW7UTG`t+&|UWDpDstsSEM()}T zUZM?6;wcsoS6iKKxf*T0M*D2)?4QAuk!r>i6%AcsM4;ANsm53g*rA?3bV!kUNjxls zGP8zZxul^bKqwRWocN*G8URZ8p&$N?$8$sm*#IfbfoWA?UCCHWidHE+o^e(RaXK$m zKnoK&@AURBAk2Ld0HGMKJxbnaP4Clq{$5S)|IG)&<+I8blFwNBA2NR0`sg?7(6oQsr+(Sp>H% z)W68ti%yO1XN~zphj~L-!+gPE^hc<2Mrppb3lG%ma>@ZPn8R&7X2N@;kRyL@07T~q ze%20=ew&R-*_B+hq=K^!@PQ>oSi?jKtAb7?4J#aJP8amR*T zH>M9=a`Pgn&X`!6#nhp0s|_Ni2&sfg$Nhn5W(Gi{?rVd_ue5!!uOOOILM`{({5VZ@ndsO(Eog2?mGIa#E$)4p zx(e;EC1|bX=+KtPY#Wz=SLL*iK62bj++5>vwQof_U3u}?mvz+3Iux&L8Pdj(ai}B8 z7>FA_)9-i24>m0Q`j6TcacEgSSvFZJnzglP6Pc?CiH1?^f;e*TR@Fu!!L!zBCv|Tt zqru1n*R}NAFE9Xki&`v*{_d2Z=uIz}o0*@vD!=^kG^P3nQ$% zZ%x3Q2zk>it+Njdvq7&amrqNv4@`n!n`RtwiX3QiqfET})0)OWgNmys6-$NUfe%Ir ztpLpW26a~l$+Lkwk|MeeDk&kSP!!Yv4Jy$`uZE~MK1r51nRIkZF}_$C%oh2x?M!y(@g)-Oq)itHohyFv~{kdW#NB68UZv_H76*~T8@_){5<8Qnj#o7 z{h8!9DK4W;Eg#sPx-gy2s8W-ii}wW~cA;V8n<>nl7m28Uq$|oQ%L>KYip>*H=2=-N zZf8IivFRK_mx#Fv7PX&=<5$tG`+nJf5W@k!_($*CJR~U~AZ&>w!2|)70>1r`n1k6P zl^5%0INnd6DZ7Yow7X2`ku9boS?dgK&$~7^LAkJN6$qk0Bx0w`~G>M?YMhA)Xdzu_gonh#=859{b38Ym@ zz?;e}6YL-L0@!-t&tw5MtTeb#dFmqK|J$KjhBt+PC_(?CxTmf}192y7@FWXwOJMJ7 zAg{k4XNUsU@G{uLJk*T-=G8ipyZi546W0H@CW_7mPVRQbbb{hSBDOY8O8>_6g8zeW z!otYJ_Fs9EV08#SCFRW?qA46;p<_-F$IC5oMM)&_MdVaJ#Na#?sQoxcRt+K6!Ox z0?hCI5;dR7pUAkMC=Zm z%xX7y^}09bAwQgMY6I8EenQQVwmiS8S2kt6QWLg7Nb}`OyII+!q;{C zU~75=M$Y1u4R9r>g#%USf5V3rgFx^1qmBaTSqbA+jqKbe#0uHf2TzUd&DGVAETS^h{vPoABuAK=sAq3#T<+V?IrTYI5}st&49xa9Hr!412pgYp7t zv99%+2MG}&ku5n%b}o4(_oBKzfD`S~}h@Swk5~Wqfw#^uE9)yX1;QNvbVky@#&HkoII&y}^}-#BSrord0&x zZF`YkcHhoj0_naEW)aT79`M$|y9k6eu{Lwp{wcRyeHJtQf@^H!$^oRe@<71%!(Eu(}~-E!}7b0h{)eQfv%dk<4R_Iq|8 zg^Mc0>jMXU`q!~vW45_6xpjw6l^AU%*ZcLtz4bIsJAIyetZ-WZ;+w8wIP!cZ&YQ(kom%4#+JwCXUbd6xH@4sK zIvXeK9^xP|Hjm0lx^jFJd*7X6vDN$Le)bD=uL0e6#wD zB#mVF`3dDa*p;^47N$Ix`ZSDY-w(hA<4P9&J8Djfc8VzgbgO8Uo34$u0PC6zY-Zes zwll`M?)tGZ%Ujm^lNypCRog(i-tK4H_BQMBfq=j9cwt(Skbu7~^#ug+dEe8?0l!x1 zg!5iCWq=7C-R``vMlO>4zl8iFgv#BSSo5&4F5H!sO-aqwz)Rr0;Zr}QR zIN9r9f~7M@F!NM=O2q8AvR(IXPK%Qw0uIiSh{i0x5+ck=^j>^R3muTFqa&HQ@LdRY zGrOeLcg!_E>*3oQW|{?uR5jEFh?*tL27@5i%@h4i!dLwfsYT^xe)vjo4~ldA?Niz* z>|Ye`;B$hT&Z45Kl2;w{a}!=Q*X$1W`APdjG0beG4y~u?FbY~~GV2&83roHDKE`LN zZ#ycr{zvzwp1)LQ0rbG327(q2Hf9q%qkiZTGlvFD?_zgHgI-ZVVdG;n1$NA!Z#a{} z8+SEzk=AN-9=hU=lflzkjsExBNSA+yjg9?UwPwUJbJ@OXy}6?#wA}zcr&SPZ0|(BT zR83Y;9ZYr+>4(L>Wt$;$i_~_HBh%#dxEfbB`-aTE#IUY*JAR5AsU=x_Q5N znIwkX^(5RBT3&^e+>G=*Zf-eL(1D&Uq?$WO6hxBj}&c5(~>(d z#`Yp=o!kkA#)XBacj>OGGryiQ{0S=VDh$0_t#`LHL)1_NC z{+5>ZhQz_k`rswk*&Le8FWFq-sn6k4nJl-~lo^e*C+KbhgGaj<$x+{>EFm7pn4xG+NCS?>z31`7hhLyy88_^@FjTXLPmlVTI zI@`RRthWllCK~rYr0Gw&p`p2T9|Pq#oC!DGnT;c+iplgPo$qlH_&iDTgqW}`ex@)s zCzgrv1B7nnXN;hb&M&@v1j_Omfi&+sGr z9-RqaDZLn0AWm8e=;`3jo8nXXOfI2@bNC`W#g@^K>Zr*Rdlx@(Y^h@1HXt}{gvf7 zF$Av;CGPG?Xq3okH7hCX}fjA-^X$>+rww!xT>W6`c>Hp4D`NwHcLLJPvfpoF1KR9Ui)!pVyzax<79` zvSdg};4YwbL`4LdUAIiAwlV@V))Eu;>gDC@V)a&%8d^EZ20m6f5+ya0H8s2%4sXDi z_AUugK)&;BpY9m=tI*f(MiS9`gx!}OSd#E7QGzXD2H zoA%sB%Ok^oh5HCSyvY=wQN_}Wr*Kc3b{hb9AUeF84Tf~AWsp6MYD?I5vr168j5weYmw!2?bwj2$=EVbrHQFu6h*&7FP*-;mR07_K#IfUO;dr*% zWj%6v*%(1>scy@UBF=hJJvYXlh->2QqEX6+g@U?@xZv2@*p8m!fNe^p z4hjNuw?UmMp?d#IBJNVef;F2Eb=s}&#+p4r^(ppHDMXKDA};c{mzGnTVcBfba4)!K zyTsB0WxM$7g~a~~3~~{*wR$pXFv(HG*T2}4377bnb9r}bY=7=L;hwp{A&QR{)kMO( zwN`Ea!u|XuTXFg^F3b>@nli{Dwi4e zIros@6e3-MmtgTt+Cs9z+;z$wG6zBqZadp1sNU(p&b>oXAdO_)`j<;CN=? zH6W%~CEr_=P*Qc3W0m7gmGVPoat6A~9;2!fEhITFZ6Km5grVj_@1$aW9b!Md%yzeA zVw<|fPQBUcn@^rfD@{VbYo8daybpdN=1V&3=qB%L=XN3Jyk zV`;G58Zn5+IDJl&vE)F{Hh9jQuj?bKXBjC?XWHb<%uGb8@T&Qy29`x9XuwX6tb_wb zj=-8Br&`Ch;FLA|_Y=ul0=(wM{%67I4?rMlaVxgj{A37i&-?hP>1rn-7s1`E`$9*N zhIJGz1qswX-Or|WmE_bZlkxpB1dMd?vN8#H&)cH5mWEiEdRSm+4Xk?Dv7f#z6( z4n{Ex5j9RiI>tb|_`zG6AgOSE@}mpz@%eH8edV!pX+_!weLzw%J{CjHB&jO^0_F$(qYG>}%Z0QOL;2 z2CVl=|CYegws$%_yLhoNp!v+2nvmxSABhRz60he1_>?6pRcv;s^9ZW*=gQTyox|&_HvO zJ@wP2Ib|wC`_ZVb7Q!2{3p3obL>CQGB${b&lTquqI%l4?3Gx-!-xzh3+DU?#SB`avc2 z*pCy~8SS5b;{JnT)oS*2A-6`G`f8x6HJbu6HaDQ@;TjTSzmn{fC(VNT`OG>iiUwt~d3$UIMR8$ETrMkg6|7D$@U8&iWPe4JsZGJ-MMU zlw^_iWIJYm4$vL2IxJZn?4OXy8$w@!!>qhfaKq4{fXzu=2KuCNR|p~f{$Q5C9+n#@ zPkG}qt#?7%r?3aYt=$q7XZH8NwCq2#_2aZKH1sykMY6M#&{OH4nK*=hM2Tp36P;me z{C2l@m?yaZBBK>Qws%FT^5zzhhQcBE@Vkq@t9aSFCr8fvv~i$&-F#i@zy>5^M^+q0`R}7&cX2iQvLq}mgZpmPr?5K zOEb~4G5rsXS4IXVR`&lo|Ez8>JtgH0das#oSC7s`&s&o8bR*JnVnGK22zY_Ad|>cy zNho~C2tP3>$Owb}fEpY-2w-}@U}Y!ORqgGQj$_AdSE9;FbuZWYs)Y1k4p&kFtIyxt zS-V^JoAz(+?k%7JZ~}~t2(u(~UJjiX*OwB_{8JLO6WX=99g;Y1HW&gcK{2P_0~SdY z+}qB5Xn_MoaEKEc^?Lhzuwx<@!XjpZrH7dX;YZXTMte&z0VdD=uZZHge5*oN0%oWg z(Vk+e)!i?HBzJ%oj<%BxluERz%-0D}$ov<@qa3Q!?Jg0!v>761PZ+CCm1)iPGk@uY zQN2%_ps3Q7D($8c?_q1uAXa94;%h->?eJR>Kw?yE#H2#>m&a?h5BsLH1w8;^2r~ep zxVb4&F-X*;3Em0L?U%?I}X$i z0|td$@Lh&Kv2|VoDO>35ub@ZMG}FIu+EAl-)U?68ydJREgN#iW+cB8(7#@(_(KdXR zb3#to#`9=zt7gasdu5q&==NoJ7zv)sA%->8ZLVXBS zQ+0E0XLN%asrLWEm_c*DnmqmK+HO39?(p;ZWB35jatQGT@TF?fomb_nd1Klh`w;T= zONhdKiQI%-EucAv+c#}^#`2`O6xq`cjcUl^6D=XvTbw*)2xy3V?0U~0TewfJ!B(hk z#`ZNcfMKhL2D|Ujh1)AjL4nd7&=Jik;VpBUFFWUJ@RL^9=Y*byBeh)Q|Qq`l#v;yd(Bv>Orv`{JcMWV_Pz_;Aqa!5H#G7@jS=x&a`I~P0wZMMX^G>ng(&O zepe7W6MOYj0W<3hbnz$SO31a)bMKY9OOTB-m=E+2T_AiNRnH)St?-e_Nc-)P=G zy#8CNFP5K>1vw2$8D#~SC2|XHmT(O|nt~NM%`&j^&~no@P2k0WNI|* zkX=iGbi4MuwmQu>!+kQsd~#w;nhbmo5hsNtFd2>**gvVT??4fru%YGYcS+1X*_^?* z!@T)2jNVCzt^3l<1rllKb8b)_Lf84hoYB#J@{td40CZNv({F%xbBGmz zA5f0TCjJ7yV!xngCx~}Gkh^qW3>m`0{u?3rgJXV-ezS1%B3A_x^UzQ5-B|S3A|}ZC zD&Tw}hrRP;^9UC{7I4m#zo_K-v-~&HdLsJ37l06iW8u7lyP>9@nI9m{V7VqiZxG!7 zbOez||Gz$AUQ2~$2;M{{Wz|c&>2PzlWzZ~vG5fs?;u#`He`Gi1v3{T<=NY_spJ7}I z1Kp^)zDM7fE)1J+zj02>eLQ%Zk$=H|;jzldarw#QVm^p+;A9U4eNoWs(<*@fA_8h_ z7sSj(-VdIk1+1R!GSNI?4DG4fkvJ#2$iuV+*G#+;^xk1<#D6OH;IhcQe*uWB_+p5Z z0)3PE#{bJfXE5yH`rhtLLcq&Yg)BM!P1&SglUXhp4v-ZdoD-?QB_8U`i;O7*?~zj1 z+sd9(gds9Ey7{@`vLI|w)ZL?-|K~;9g~II=p%9PSHv#j4jY9IVE&HRsrEo$jGzRw| z98F<3A(VZq#9eDx0Gy>4yCD|vNz!-yWd~lm>z!h3La2&`I5VCr5e0An&zZ45E5&~; z0i^qZy#=Ykk~U5|H)A|o*o+4wuyLbORpp}ARLm}~kSd*+ObB-~7pKOA2n_(}oxQa0 z%!2KlwqVGN0_~Qmly9Ug7NAX(7$VH~%cqDkh~9~J)YkTI*Ouzz@^FG!f0+r4aTnw~ zacR%9JbE5=YCm}hZLi@?2qZ%LU5wY-7qht&o9bvF*F=7f?Ch)YQKmeAUa+`s)rWWb$$H=OZn-`}B0Jm&>&&uqQ-VJjE(P8F$*cJTm?zxPuRb9unHAy|K zR&O6yFO<8eQ+dG`L0L|(bZT9XaVQ$W8M?c65m`ZNvI{`{eTT+t9h-YsYv;HB1e_slrdZ zBbVqv^(}FrS;Syq+$dSCf?K)pj>rO&YkQ^S0nWbmM!Og7-r&M|FsZduZj)MD<9lpe z@%-{S=l11IMWe^&3Co(yncQhn^tQ>2agUFz_!V#^ZPW=waRvAoilD!6;HO&{Zx-Kl zB4{X5BH05g0E$yPYTKT8iRZp>#h8UW%7vC3NNB$S+xOUhDK>S@)^Os8paZ7&a zN2M#wWdLXhoOBv&X|wXE{?xWjz;ZVab_N`lwkDYjomJ)GiyQrV8ttvY*PZN`;==%LJi1u?ZGTO?_SM#W zuyf|p_0O-Hg6t2J1M038SM^((_VWB?t7Xk*&&~Z&HZwOpQeK9psgsLD?Is)Ikc{c6 zSgobe>DG)p&wKS_H|@%%JE~P*%?wMlBN+WgB5&*J2Ft??t6O++8)jLu0FjK_{Y^L- zdza@!)VZm8otJB?46nF;h0je`OABH_DaFAfgcYCxLAcyD1*9(#^0Qnd2Mr?q?wlB< zK~Mh}H7aCJ{uwTJiRMH1Zf3!i>iW+W`-?mC>ACuYv%5=NRZ42J2x@dk??Wf?% ztUp$KXpR8;U#9KcCjAWy`Rrt(oZH8oBs&E|63-l6xC~3@X!6xlRl4uR6p&~(?_I5M z!<$?mYpn?o^1{Z#oG8%10sc-E1zpfV-Eu{?F`q$F0jSGC1J15DZ-rzi50**?K|3AjT#BGaD*kNpk#@K{(<{*}ify4Pxd?J5CdYprQg={W(1t$~A*dZx6(R?|cSG!b zMq$})nV?6R`xT6YZ5>hLL1c?l>`o*hMh~XoTz;jCl!-P>KR<4!w zmEmxXjHp}q@!T9(&p9ZAeuJdu$upxx8s%RAV|Cp!2n9vu=ru8o)-_oe(P7xzg2fqG2eZ-|c{zor{RPuKUd5n$CCKEH07A3_I4 zF0<7jDvFNs?{C13I5V|5Ba+5BgwfD-fWsy+U?YFtpDBiu_B|$2~Y3H!@`@3C8A}XOn@vcB;;KYej z%q1$=1*m&IqIS+93e>!hLRJmAF8KyPT-*ojCtZyHfe4)xuYwQYlDogeUNdr`dT+Sr zAmRUI)C8$Nv4-Pna4PA<)i%gmU`jBsa;;j^CkMEzeJ<$yt?F_4@xh$#rfbWEQ5gff zHIF-dZt0}R&=(Ebfa<{`49mh<_Q2>M_s8_dWpOFEa7F5|G$*Tdzc0r$mA3b@b7nI{ zbn+V&s?)v5cO?2SRt|6Tn=MM4)n{C1{Uh^6gW*!l`zbtTdiOe21K9~BiE;NYeD1I;j|XNNR491(OMMwi7;yw5Xb9vd_Y3tkX?Z2K1hHbCw@oHS z2mwU@yinNel)hq;1V{Q01IXMIsy?;k?9xFoz|};1#(YL%!(fqog^$9R33m^%KXun_Xi&?P(WvEx{Gn>7;JXh=SpD@qmw z`}_c2cTP%y+}Kbnm0{urEGF2vh9iZwd6UmXX9yq1Ez&KU z#+uWg*2Mr51+i2)^w`c>1)ajBFm&s*7cd5u^0o~NFjASA)UG17OPc6t(ggZ& zK5d)F35U2{N6?6BKLo04o^_dt!z9j#5+wE#C}NoIJbV+FO6van>=+Z1 z%-rKp=}YrlR9LFaR_8O_jt`!*Jpov;bkp;ns!>|6gU^Xw&H=^!azy>mz-rURtF z@}8u)0RVU6sx)yqE? zE&Gx4iYM!|pL#?`F;ic=@i(&liHPGf^@4YS$wLz3E;T=ob3^l^moiG=S5+8N9&H)S-4pmBCe@mo zLrDoVL=7JWgaTY**$f^cm^(R+w!+h9lg%{eYrHYHY5DukzjJo$^Ccp4rMb1bLG%g? zZ=>1bnuw{qHZ>>x{b+Z$tMITjQsoA#{cwP(QZ?J)d$wB3>oB9`L(0Ol&iqjTo~mnA%XX-IB~)&UtzqzaLZ4TQ8px|n|ZgQ?D=$wGLjoogfG)i z@SgBQO*aq^@W^(&6)WImyqL2zyMm&i>x$4@O~#L!G_jT-Cjv;k_Rti_>~`SS8?Fsj z^}lZ=)H#e>u6l+v-pzhf{TTB)b6@kabFzKPbHsD|-FQirDz%5|2GSGMxs`awvVp~; z$W3((#S`g`<;3QZ^Z2FkOZ8gYef!x2AS;kMM|LfRK3rw5xitgK;pj}+XI`nO(W_ec z=-#cD=%%F*Jw@Bxxy_T5jCE+wb>#wqx=F&?y=?P%@vPG-dg|iYg;}d}NS>zow|C}t z*@6UDXw#aNVhGeMc0v`}$l|i4EA!`05P9MF z8?*0TR^3%O*(6$onsnRvo#XY1-MJ+BpNbB{Oc`t}NpivExLX6_Lp?wT4|$O*uW}n` z%TUsYLL<2@Sw${^et7nbZJAMz6xyKL@-K6m?9tI!%UKQf9aH*PEi{lc8(R1`W=+S7 zKjdO{I_OJ3I~`WLr&e^fMSCt%o`J>8Yh-%EC~XZ5T4QcJe+lQk^|J4@1ISR{(fw_o zDn4$qMNg&MmKtB#uJ4uGsZMjp3$uNPLw)P_QXmt@(Ga%cEH#p|x6!kD+h2?Ey6)Gj zH)D9aUoUv@4oZ8%E}iTVS*L%oy~t6<9wolGU!hI^8ke}F!DZyKE_PaR1-@?LlDn^b zSdF`%G@UeE4pOu6wapcXn0o-M=0xMmaCzL8%P{u*J~TS z0;H#wP*UlJZ%PIiQds2l(KhYqGyCY*Yo?>eycfQp8>oZDbU%OM8G8b3TPta|M7FU5uz$L*;8`iBUgi6w;%8{f%I|Gj< ziz>|)hORj>i}-ZhCy5*n{&mQbGL?E~cQ&C~AGw@S6893(O@6}G39ST^N&yVu?DPH` z>D9-=>Wq`s+03O!-9&)2n?vDmJ(EO0NxBiD>GA1VtJUEUE_Q4-nOr$ClHB>0ag}72E8%4S*~O2A3gz;#P`W&@8 z>&LuO*XSNvwhKS+lQ+GKw6ro>Sn)(_+b&kP^ol!U^SJZ~Q`3!0y9FO5?8DJiGTU6Y zTzoL};E@+3!vZKH;&8t788^9uRXoI(YfRW*q1Frk&T7y2r0HYW51c9)9B9Pf02xw% z0km`82Ano}^fdUYA0T4B(#eG!RlMnv59%fj<{n z--Wj8!3DemIeb7j-=7ykfVh-I0FX6*z$-pE!LmQ(d4(Wxy9{ApSH1#&F$2zhhren2 zzE2_vSvBj7Y6RT@4p-j?xe1W%rwc#k4#{kL%4zLX#aQg0%VBCJu#Ud{3?;=ZpLJU_ zCq^eQHV6EwD3CU&^ge#MO$=ocVPGp~Y1GdjL&G|$RW}DU&yNBqBXtjFd33Qb=LaPC zU2Z0QiQT}AOhG)du_{6KPZW)u(v1`yLQLu%yN$XH42=YzVBf!In9iDxvVK|@UIi^G zH%(eKm`yrK3HQ3N|T|!5vk(Kh37Tu zv9d$CB~jXZHQANXZ!k$4Sh~s_=Bahc5Yr)F9#9P6UY3n{Z+n+!JmH5N-lWMr&`9#lA4T*)N$0>PTOBz9d)hSZ-2LTzDIc2Mqo^8VT|sROm@G(9D08pZ+vye zy=&|bUgqU@{8h@*d9^O}~6;3tc2+FK5)w#+|W>um+qeEu?vZw`tLWxd%f%N%Zt8 z`w(Uhs zi<3lm&cxBTb0%Y+1k``gB;qhJCG`}4cQmicgAaaYQ4aZ|)BttYrq5A1jFXl-4?sg> zJEBHsYw;HbVwF8yy#|+asbfIi=CH~l_FTN!*=+UC^+esUw0r24xwTGXtN}Gbz1Vm# z&9?n-d4hQ=qWtGr!J4A8DA8Dk-{M*}C2XDbB?LAp!a*&iG&(^u<8dmBtF)~dGjR~h zWeDRfmXL#o5lzJ?^`+?m2r~S^i<43mEyH|KQmx{zQo)J+oFVi+jhLk~&?XlVY2l?1 z(wIUPG)d5iZ^i#6cmaq4(ulFk86g%>Wg&*k9~@nSUc+|+G%>}J7cjAR2Q7|W5NFB2 zQUwC4h!|%Z`7x6OF^{XXN=ja|rLbdUx{t~^J8}^glazPzS zqIDcwY1fs`CA0yqPE=Teh`}>6knR?~h<=}b7f8=sAAlV@jj0}Mo)m?r8VB%2_%#SN z;SMRy#Q+QB4;}*moG@rCYcs}4tT?0(p+LH8Rqe6nmx9F}6%FKe;v0922zPp%#b-%T z%;V7GC^}R`C2Xl0SIR_(>h*FQF?I+nT&a6VA)E^v=0Kk&yKJpa!ZN%ySmkm)${KF| z?c1FQP$7%Q0+&M$kt3$&Y#An}wf$aINDl6YRWKwEv=X;GuQyWv&b`s_wREf?Hna!R zDuyb7c_oJHXnX`V{BAq9`<*}CwApbTT)PU2$eTovyWpwEXs!10FW|KD4}u?gN4X0< zq@N+6-k%`RPz+m;Vz`34g1Ld=0SRU`tt!w1PxyENbSx(k?m?7vGN>QAgnZT`ICV)j z&|{#61?dh84;v&1PZORkgvW9SKARg3dpF)=gP?mF?xWWaUn7ToEciTN@!1E-Xyf}d z(1A-6o7`MEs0P{0Kte|f}-;zgs(Et2#_6n_mLclLT?mJjr&hUK5%~CRb5-t`8 zo)kBOYV~9PaAs{J15diffG_CM*S?dQJTFvjo`6^hn>p{kD#NZ!nJBMo^Ue}R`kAs& zHf_DNt(IOFQGc@v?0z!gY-C3W)9c+uYuRyrR&Dem4V*Bow%bxtn7gJEoIVif6HaJw z&ZLNfXgB6JkQ!#fP8uPS*%Vb5-6-8GSjF0OxItmL5>Wi2ax_tJQz}@vyaDLbyKRF+h5%#lGhs8>G;B@*rA`A&2IX|WL;)qH28H1F zJX=B?gf8`h+iCiTM5w)GyixchcoLJg9feq zqGyv4vhnuoz0G>Ih#&yMZwR=x$x|m_By5%le7})@GF^=ZE2be3J8T94PyMnG7?Dk; z2;Dv+VKHl>5t*sVC${yR-H(}OlRE9kysyFa+ISbcV5tF>l_gL_UlV(8aL_5FxGa8tqx6r*Ve* zo46$ZiZRUbO68N@KWMUqi=t@>Ctmy_CsxAX_r@=hiBoQjA1J_>KBx!GVwet`sCorbRj2txALV$}) z7Yq$TVMH!LAUIFW?B^@`A!ErH!f$CeH*5Ls#7At*D9lzHEB0+^f95L)BJ^S^JYj%N z9l^Olpe|aF0e=JmYtV2O3LKzuV1ulvl2!=K>w6oswG3vm+ej!^ueR}9Z}omOB$_|f z#m0rH>O$A?+yXiWSZb^LupuIAvx|eEDs=YB#v{$-dAOs%{|ohkQF=nNk#}MYrx4JS z4e6r=lQVF=hOmiRYC>f~Wm=g+r9!1zsnVE;+bBK$GOAK#Ck{^)R~5%Fm)85Q4{9%=ulPIVrnU#E zPw|pIc{X__er{Eo`U2*^KT>2H{oU*s{Ut^rzd}fUB22&qyx4wWPN4g>VY9ll{ILxg zml-JALf)M+nAl2*i09H3Dzfw_ zuB+{>9ph1=;jlKXul`ZQ4gdnKT{U2N@UEiQmJE37%odYihZn?ejBz^5al3N5IYRp|Wobdg zHJzeGT89WYy_TOgreE10EepDT6_ro2lxUPHt`Ri#%0u0~J{LC04$q#FZ>^?iwXvF( z+Opp+ONMd`WvMO8GQJvv$HT2ZcTymDL?T>>9h=pEM+7rTqqY)050RKszx3J(Dl-1O z?ZPo+Fz#ux#W?6>pjSfb^)EtBL3u_)A9eZdNVmImQlLUH7<~-KJ=!B7RL!)hkHuFJ z7>Y)6tV~Gwb3sA!86*ZA{(zSXCVd*yw-#wcuN^?BtFO zj5T`C7Yt)nPwJ42zgW5WNs7PJXB!dFp)STN%Y(N)*MFz`)(GLRXfR})JU@^?2nsQP zN+nP$Mj_Pv%UR-f-Hz*2n31MVe$7o%9}&K~2WOS-vj26;)TL*A-Fw?0?Y+K;sw!?-Ynq&YKCp3Ph`hwX#(K`LMbRRtlC z{t67x`@;ZoWD3|9D0yjo@+tuDix9;h#1{`9DFhMa@&cHvhad5F?7s)AC*2)RcY=}t zkbXfe1o-PLNa>5YDufCX^O^^Ljt)eB9_}K{gp|8@DNo2-A%A|}@o6-=nwYTsnpze_ zDQ*jNv;YrDEmLyaJpx*r>UDMctp94vZA@1cKx+}wdi_*T=GKZC)jiKcAX03H-U%@1 zWkqcpkUi8^ZNSG1v2O{zK^_&%uc$p)qEC@e~k6JV)1N#(S#@%(HRu$=!@aTK0X(e~?Dbmx zTQ57Q+wHP@sE_p8!`?n03CA{~Zo|VByt~L}R3O%0;Q-~MC{zF~aK@tRU_$sIMcQ9sao zUkK{<#0U{+Arn=oR$CSgtSZr#b}a?$8rU7l8`nEXywbeKTsODB8ssYangpD-kX*up zxPaZ1>zF>Hve>_$xvNs{DK_vP0Dl>LcwBe|f`Npg+MB^?$__c|N2Ioi=~clO@0p{? zKS8aq-Njl)aMXZ;_98h(Jx1|?3KTie(s~gd5O+Aj>nh|p^jX9|XiVCn@z7l%=plEI zphC_ZTrCL~mjM;OdM)uyK0YSwSRgQupCI4r9)mA8gg8gwu@s=O9C5jQ6!3!-w&C<^ zkuT9xf(`mz#?v;dN+7a7r4trci~4r!?(YjfGd5u!EO{PocHT)vH=^)pa#~035iqq+)W!|NV zXy?R&3UjH;FU#;>{cXZGrVH4cg#4C#Y3I_US_m-44{z)V8Hb8g+?hNUSmL5#&Kh_C zZSB%UPp^PL7Z6cQjtX!YM8ZBiM~rx0^t}u&j)r z=#M&LH7%OOqm1l2=0iC_8S7h)EUlsi4>C{j-3bq)_Ab z+(HqVW3oxJOPL22_lQ6J6Yr_zjpS)eK4}A!odyBzgMg3XHK4Z~4WcxN8cEu@ZpG4M zWN8aj^$u3sqFVLz)lvkAVCJ-{E=5m0$bm4%;HnsDfl3D}OQGDvKqy>tzdFI^r@$%K z=3~5G@3TogQ6=xD`Y@6bax2|ubH%=fq(FTF2L&R76(!`|eQiWDk|*s3AEd1yadMK_ z4FEs(Pvnj0FV3mTf>9rFht2po?&h;sAHTg09Q#(c*JHT(Y)z-T?ki1uabg7jDAx40 zb%r(iI{&4HFQQs6LyOt-I>^v4noH+H!Iwod zqcgS@Y`?FO4-Yv75BHQx@LI4!uK1wopY$F+E}RCV_m5nS?(!8RST3syf55{xCb7BJ+t(?ih{u`vNXb)!{~@7Hbw6NAnJAxa}l z4G=?qRZ&dfrh4{}GF{=EtV&!y0|S;Dc>%-fl@ejzx9{pGw~KpNY`;4R#GsI^2@!#| zkvx1<0h$F|oi$3c#bK?+c1ZK4{pF=bvLJRA<47Iu`}22lf|$yf71cNohu2N5at^au z0_C)`@w3^LWrvYyE$8o$v3~~BHhAE6``(`wwp%4%F-^8iNKsx6cW*9fo6#{>qDTvxjD=pN16h)677 zvQ+VEY}sK9M?yt~tFYvvpqY!(8AByt+nc!jI2lO#ckU2bznsfhvC*4;w?HtOZ0Is~`h6>>2qJV>|?kCr`W55>*zK zAG0hSGF{Tnp$5V(#l~eDshyM}FUAf+A$e(-O~Tr6AR|h|n6YF5N~%XvOa&=9%Vc#jVM@FXe>foH_R*ccHETW;MHJ8RDs`KMOB)7a6k$oZc$Lu1Om9)Q4neLm& zIkw4)w8|CCs&aFqmU#Z#U9$#Z_a?&GmhQuoJMr1wishqq=+|9Yqgl$tRN~XKE%B(- ztWoP6wqDXqxA8kMcXn}?GGXIdXLn+9-;mBvwP^_naF2{KEWAIEs#FXDxkea&+Y203 zT{r6sxoRVwvRqP3O16&IuSdX)+ZT9=QKUk&k1b;B!~H^<#zhG^m%FNS;M!&n9JEd2 zB*t$HVsjCkYxql(Y6?$yJ(eZC=uo)4^kM4eIXJ901?Satlbui&&i0jk0_4HZTu@t^ zwS=h-*6#{%P9uiSBP8rMJ#)!*5{7f81)H0LmHVS~V3=a#to`h;BUSW_qTf#&c!%4C z^j{i)&rhHup%GDu!3{w{R^bND0~}c+7FqnVQ_RDx ztEl&uEm5g_%Qm+dsteh2nD%a+v|wX_gk1~Onz-8j^!cSEtu&A($4ZNr64EfC>T8%X zubT)W)hz?se`9(y;pCqpIaRnyI76vY&0H*5#>q;XDs@PjrO69%FUvO-Ng3;Ul{i)< zHhj(iN<-?Z%*f$@TyY$OoB^aj!JTI$bkivA1|{fVBzc+i!??QYzg9x)8=5dZ%uPXe zBoL_34wavAnoWvary#MSVQ-V zUquj7b;;l^SFQ-+gkVnvW8^S+WI1`{RM6qXQ5_FDNL+GQKzf-hFjUu<;O5Ta`Mkzd zuB{bc!FCG?58%0?Lj@+|jGW1*eG%!zxQ0Os5$e?ZG9fFY)L|DdUDk7;usxx3;4L4G z3fg{Td^OJ}I3&^yYo zbWpWwUUN*xCKA*H(pTr^knTij*{0|IFfb{l z(*Wvo$|Nu{c^CVMji=SO-K!qCZB{N^v0(K;@zFzFuwJ(Ss|y77d$yOrU+#YWlC02K z?kp)_MW{I{g^L)ia2B&i^`b4%sAgCKxb8n$y9EOF)51)iexvX`h4U4!m-(wOJtVw& zf#;Tbd?LNjrWCPrRg^eB+er%!;nRF7fKRVFpWW{{AV3DbfR$IF7oJmSOcN$2d`eA2 zCHun}6jlCjcAI!`Ms|Al;TqK-gei_wnF>kpNnMQ-xR4k}ha|J+3l%O`k1A(#u7t<; z)p_yNN?AF+!jMGvpGX7Lb*XSh zb+Hezp|bUv4c_P7_1){dPw|WJRCskBtsfe`DvzSakE)~0qm?o1-hAH0dy}J!D&`e) zvA^~e8d@FQfq5aDSY)zAZz|Vz9`%WODH|$|#xJz6*kb&dh+*~o$T+C^jh(9duQIc& z$a!|KvQjd3f>Bn8c36KB)Qp#Nq^eeil8F|M*%Kwq*|>)w4Gu$^J2y3ztE($3L*<9a z~VaQ zVapw|&xFoM#Oxi%Jg8W?E}f87O>ASatzu-y9IRTu4kF+Ri-_X;XspZz@ouc`0)7c! zhYPmlmWYGfLJSBnJ!dg4KIER}fvsq@Eky zclfyUztcv3VIQR%xS)k3Rs@qT*{347lS5I_k0yK6)F<0leS}$bYAoQyp`sYZq$eA& z=!YRW8^Miz2Xn$0PVDYdFZj1a{KWFY9Cx9iN2d48Qhn2lF-BUr+;7Hzv%lW;X-I>X za!9xfSfab<{2bIY9=ytWsS&A{U*+5+S4)2}&UpokT0_TqcGKR{bp6CkStoJ>2@zDl zu(`!bn4_PXu{B855;3WE5hK$;lGqJ{XAyHXRUYqA#@v||>++vp>?--2Z+ z_P5p2I}D8|Xp^Tb%3Jo?du@9uIWOKvNpdVLo9!0%EV(HBt+=p^tq{_-#VfUCl1G+b zX|@to>d?7KjvLcKBgHG0LWuTEsmb+KdSHji{C`j7(=UHX6b-99lNPo}-7F_Auw-K>-~A`KShw zm#&ilDPnek)gWC&UbANxj}Kletu?^~CtrJI-C)Q{Ubj`iq!s6}>SPhG^K*E4U>^bV zNj5*@kj6xo3o3ClD0e>4gWWOWNde`T*ZI2lOF9c_&5))vkXCWLc^Vcf1^!u)7ad_W z4^-h^UL2``@+)>cS!iwDati;}T1~QG!ceA7($U!D+g`JG>AbE~ZJAP{PLsjfP0@LX zCh0jt@;Tjp_^UH&Fm*DD1uNIGH>YA-s4lariCbjQ=O**A8Ej9V!Elmo4#lPq5@w09 z&mHqU09!5~*J+p;Q9j>9+Eki6D1L5SMO;=+n}nMdoW=vkt0Bh zaamO8fZKHh1;be}lp_w_0Yw8f&}=w<;nGofiBRU57`9EvMtDGiOwxReGeS5IO+IAb z0l_B~i;13&mFDEw&>a0Va2N=xBY53AH@$S*h>Mm=8dJnwQJ%~%y=msm|!O$?-;^(D6h!zDOwD#{FP;%IPCt&NAuh-wG zgL|Jb6=Mg66lrRiO4{r3mljLp!t1gM%G&)&v${(nA>|=_7&6^N_G=-5WS*ETE4oaA*{(>Mo=7{T4 z&Iay7A))5DxX7+3L~#(4$ z^VblBz8z!j=%P`dm>7P$^b$r3p|VB|)U*RIo`NT!W!?f|Si!AI|t)DRenED+xQ zZy>qXhlYD=TV1`~8D&-px`llK#w-&WfAjI!WHl{2$`zkfrO^U-;`?%b_y?%8wpI4^ z#MW4^!`S(ToowU+|1`Z$mw^qo^%Y&8b<0$dPEp*ZV8V~dSy@;Jw|2vwo!td8o+qW9 z$9*eh3!(gL`qa{s8`L>LWxstWqfWm~u$Se5CJ6xR9-tc1`-MCto$?;#1?8&TGisUd zs}+l~WVMbPdJCvTq!U3Y6f;mF#D*;#b?f@|ZvX;*|s_+pf|%I`QMHZ^moY~ofkxr(WQYOM)gLLE0_z-G!pGA+Yh4UDQ7wVDv-uKhw!WqM8z#B1gdyq$Y^<K4(7tvkpezOr$f*MvXcVEJ!ue zDp*nB&zu#o;8p$N<73eX3aY8$AQcvpr>4ZQDWR?T3LJ1YWwmK(Zc+z}BW%qEdUJtf zOU^{jq7cP!C2n1A@lG9Y4XVXJ^!_md(V3KE<({j8G3~mHTCdLFwE0!1>!%R5^aHI4 zyHI6FY7-xvB+(N?PL!af0D~ZfA|5EEs$$;)uEi4e38v2?bp>E)AT)5AU`j9<{V817eOw2> zo>)>8P-?0=_7RChR^Nxef}t|FdDj^H9gRE~%Jnnie0lnlBV`@SO1>sBO&zJ#_hZ6M z!Dc<)ACay@urA_14Y+!-n;nDg9b^bkUR9$$$$y{Ibs9embM$OOwca|yUQ!p@`;6>w zPq@3ajw1NK?%9o@JITSCmnYWVea2w@5swf=>Vz(VKAS4b#N@v!kbJBHpPFu0L!EJ3>fj&>{W}KLw!xA*CDKODV>vP)M>vr z^&%t)*=+IS`)5p(qKEvfGX?eVjXn~+gyF{BqSm6(vEdRvTbt05?oEtiz44ZDYujg8 z)3K=|#l9A2dEw!)y&~t{#*FKvK?{qhu4(M)>2rPO6~kZ5#-~Mx8AHo5)Gq#DB@uEp zqb608u2Bi`kPze~dc!hxJWe#VJuZ$WD1zT+23uU~T+>8UT|=ESRU{8CFKUX#cbHc} zH+&@hA;1HpA~|q!>dpnl-4U{o)+n+&xE{ij`^{b)2!Ihrl)VDACpAHlg$My^)x>Jm zz-j8JnH!S^h$jzir#|39&g()-K8ByoE!(DDlzVPW5wbNVS)0rDe{xlLL{5yq z_3n}XXIAY!WMtGlNQ!zt+Sr}T@W=+)ruhbEpH)mfmx7=&f|z0FlaLrh+Mm0iqTl;R z}Q~bpMSCDl2XIRe){22ir?K(v|;p|9S$rA^SX{tc>L*TU_ z0ykzJ&1!Yrd{~i56vTmGA$D!LzX-O16-KoPA*(d3#16U;dHfm_IF#WhsYDYw?^YLs zVWaqjDq9|2E@VQN+cjymXnu<-c(>CtHIS;!Z>TL;>3utuGo0&O32aLQ99xqCbEKfHa1oD-$AO#E2=T|*~$`XZwvZz z13oz?;a<5A100C;uH<-fdh!d%iCW3YR2h;El3bEQR%QadE$Espxfqi5eObSXip!rI z33Z)}lEC~5KvZ)fz!YY5=w2t3y3kC4cct{8f}YV(2FvM|ySRQC+OXk*94Z9}S5|MD zC=EB;lVOtS2W4cLccT_+r@Mc*fxwjJJ{|()2-YJ&i3+M4EN#tFnyf^*gr?xHTH2r{X`H?y~86wmFCm`w5*njVTJq+XRQSd$<{cl zuG5X0GQ#s+Ub|&B*F@R+JH%LIy|z>b+K@~5!#jl+ z<~7sq;)OouDv0lEpMDtdkQ!9yT?(kpf)<=q$=(@)A8iKQGNuN0CP~}v&=&xB@N*}mV*^Q{#yJ1@xTEIX zGImIVwIum*u!G+o2A$uI%^%t%zt;AsJlwy*n>g`{%NF6@bB%$fy{El_X6Ezf!TRnX z`^de-edHR|J|nM_{w4fNF#7Yz`l!q+=67v;dGd2n?kU%$fxbx3!VyEMJbumWZ~8rX z+4McqyY;|p-hUFX7k*ot2Sx9bbHWZ95-81M+cM-cp_wtI3P92?v@?c+(FR^|!QjCl z51%FTH^q;S2Mxc{i83c5{y)=p;q13*ks78oi zls)yJQA)umnjz$-M~)5zEuG9Vvu+ggBaDoTol7JKk0h~#8QQnLnVapN!z?) z2&e?945}d^6XaENJp?V{EX|fZDa zQ14GBk<6$3Lw!O_{JVb?Lrmm*9S;rB7m_q$n4e@o_=wL>$!~%A6GOvgnlx85LrmYX z6k^y>X%-R85xpylA%}Wr9mLSP^c<=ohW`b)D8{Hh8QmbmcjMcCYIQV14F8?DKh~#o z7gH&O9N?-_;-v}p$+7l9FjJIn_76D;DOM4SW(Y#69$==-)2n@z{yr%A#D;G={`}1t zhh8=i;H^KtD@8M&>5mO@J>6dg5X0U~@sA97{r_!=X5i?1=mTG8Ba(gXJu*b{QnLmU!K^KyjM5Ii z&Cw5*_`|}dCtO+JQUAUV%O7Us_mA}s-O$O09{%aBbOpume616L<}7|Kj{3J24gSev z!41u@Kl&BHXW=sF2QzTmO&?H!dV0AJViPTc)(e&9%24l|)kmJ!rWQ1#=1lO1={rJc zof)S#ARqb^MIY?PeL!iwEpAT>>qs?~WbBcr6o>y(5lHej#Q=$@-ajFtyrS{H;HKC5t7$4Q@z23?jPOj@pn(xQ~|TJWwBG z9vdrxN5)#VV6caLm|3DZ3}KfY56K4_%nj~-|JJ;9glTK0 zb9@4sXlS$B1omXU~}`PySo zH=C;az--9gbY!vYO<_1V8Hiocd*^!dsft zxN1j$dp;A=((sqVPKHJmSRMi}ahHvm4{M>@8)xizt4bH=DefRxSjJhwpEV{Wu)j`Q zM-D>I13ozd0|+JD<#qkWcS8rgumbrC%b~&0J)P8VsB~RHWrh$ zh$1_D5tCx;mz-cEPG(T`1`GkFk8S zYhZH?l$<+}n9oIJi9B*?psxX|T2JVlKy8q~!2*$=K%GgGDGL#xGw=?EouoQ&` z51v=pQn?LY?=mo%F&R%f-mo~+F#vBGI1R+qcQSu4pIPFFOXq%^7RxJMLEa9TPRxqu zQ_VN?NMd8_1?m8*`N(!t=?#8?DIX3iIX%_8{~5m3iQMTx>T)J^JCnNpxs`$upOKo? zeOd@TLkxZuG$TDr*39gXuNj+PRJ+%-JmUTkw#by8oY}vYkHW-G%X235ngzys-G2NU znb&{l=JvkD1ElJ_X6}hK>;3fowlScYYHG3=GaJJbpGS0R=13_eIWruh?kvlU|Bm7t zbZ=m`K3x9KQg(6zw==HKC25BEpB5lH(H0U5<85h4x$$>EvGWLK7xAcucV{2Gu&obT z$9iIqyo=I>WfrYHvL3O}MLc0iqm@GPwf@Mtxl#_`RPYNdS8bHZAR0Hxb7E`J>z&LN z#}s!*2i4svN>czSnZ*zr6oTBRxXOb{p47y}{TrP{^EPg5mGh);)|z(2FD5`w%x};tZ7GaBBvFqV4q%dE zmqI%6O)=77pXdpHWSnG)3Z_P;*TCnY*F~RuDt}imc=@|MMaQQ>6w4(RozUs1g7WzX zN)yd1dC$W3Qhl8kbo112-8W6XDNa9X?g2YfaLFF}NUa!aIuImcUz0er;*Kxd?q%~? zFSFvLVIg~Y?azR`X3AUW!<3Q2@LgcXMW1Zz=VpraVbk?YOK1C((){20NuXd2_%XnD zi$!dOjjv&wvGX-YdK#iOWnc8GYIP~hyLn+L$YOXrB@LaZ314s&x2A@c^qs9w8HZH+ ze5KPJ=NN&sA|}jgHCp5OQ?}A;DwI~o%jdmd>~34D4)-x(soy)f5)^Oi6{MK_>~Hz{ zP3`iu=GE#+_-^zbdr$uxR}@e8pz3CdY^}DI*5;^_-B^WpvDmhetM)LJbPs1ebaa~3 zRjFx@3r9%}V%Ca`6@7BjtXIv~Q<4BSi5cHn$MX4bn{=TS(tUN^Ds@eYxRvx0s?y0L z{^>FnD_iAA6~^)ytyHGwuZtO}zABYLOq%W_)ix?q8D)V|GBwJJw@SEGpZzP@n88Gm)~XYl3f~E|mdUjO3RS)q z&?A2=otvTkY9CbXRP;Z_6nZ$O_Os!Upe1IN?Y-?i8a(td(Ci5~%Z_?D==KqJQE1>8uYEv;3Iq?DeyN4e8aOh!oGMXyeJ@z0y}X=3f4H*5b@T9VPZap`thB)t(zNE zWO92zaNK+z3rMsF!re_C!qN`T%<=i#K@!@L0;xQHUI%PMJ&@rJP6P8;2a*H9_>?<7 z*ef$C>LBdw9r{*I{A+dZ>buw+l5Mo;Cyb|LZHop}Rb7?4 zc-u5x{E0hjrf=>hZxPquVb^h!w|pBrlwHDhe(Z~O9ggHs)z}rggI3>7?5lVhQ?;D~ z{$W2j+vDflQ4G>a^#hNDRfP_1C)VJOnjS#B&5{(SRJ!eN8Wcj9!j8w4OFg8)~sj9X5j19;C{$w(Hu@-QhXzX7Pr*3;1-ifug^;Z;AXMwa?jp=cy~ z53v*Zf_c`Feo4=Ccu0SmyUJlSmGBckLTdSg+@VT$M!wJU81a7>K9oHR9z8AE`*H^M z5q*um#XZNdp8?#@hynH95|gG=EkOROjvXt#`_N0(Yjy2YkS(wu%rqW&S!~Ks8g*8d z5|Z6cPKw*A!h1*n6}MBY27e8}QK@_q#8>Xt+g$Td+PuO}er4$;zInw-*CKt^PrR+` zrDXNCs0j!kwx(v})M%VxCfix#Aj7rn9X--y`UXYuXV}Ma9r_v@bNa`U%%MKeXnsgC3O?7PF<^1L9i0f_K6ml@_}ci(9opOK zka)4^rSmYT_7QA9e5BpD3-gXMrMHVTL2YI1Aw`=@<6a7ko#Xh0JgsJvUF&MMW0I}@vj$v2_l_Sa+);D0FVGhg2 zNbck!Cy8W;BViZIU=&{*^?-D)22fU++##B-EmFd=BNnWx2DGd(ONgwp0wAlZj_|dJ zyJ^@kw|T`hB0pWPn(4J@q7SvHOka~s(CDTw;drsf2>h*sT(5D7z@O!nNrr>*c2D!$wmCrEqynvpI7oQ zX~J~ToM}i8@qWLohV5-NbtXKyD$Y&pN8)$G(r5DhL1BL*Gk$R`1hq0N!F5B<*~f2e zFTlQ~^Uid`7v%K)Y{TDGm?X_i>xV~YipWlNSK4%I+NX8GNJdF_GDrPZRxbr@7cGr3CYsXn zr@J$5RnbO@#OB6g0O%ORP!rwRmultmDhc)apdM-*_zjLDbWeG)y(7Z`BAl)3pPyv7W6Fl8i z>V6%EVkk-7TzVmN2mm!9X&!19*2V+^%UXJZQim*bs1p)+3287%)}fURp<+t*-hc?D zf?P1jf2d>@5V0(OcG1voSkE7vNikW0e2IA*mM$2$+>=JFVV_0H=JEFjtRknRkuC8GtT8ioK?*k+I= zKyC}r!Ifh*!Vb(PIN>lb;>10i4?p08umAx(ofD0(o)gwk-OovoK|ugYd<+H$gajK( z5Fb#q4^Z3(^2REWz(G(6&jON)zJ^(eA){ehSi%Atk`_)11{`HIKg!$uW$(1 zjN9Pw!fwPu*SAJ(7IcLg0Pu@J66XteiQlW8pBU)C!C=S}OI!>l(ctjXT8=8(ri!_P zmTO|g)TbKPBZ|Rqms)agzDL8HlE z1qkc41{_DDFJ;rmuRI?f7&c}5z5lz4%iFEN|91a$dTOr9JB1m@rY);0Ut3al#>1zZ zzZrV2%g&y8ezeivxn0c;Pm(uly71o!}WUQiO_vN=mQ4Gh+L@1gRO7Q|i=e zg3=N~&D&qXBjRv7Ok)~es#Xd=G?*T-8`1JQe613mi$FReRj5%3$VwX7!@($>+mw=* zsID|usB7#s3C&_0X8B$xKFpW!%n3v!w1mby_FsvC~ z&R7cS7WzU8pNu=gmfq3En{OU;fUzvI{qp}B2{j- z5&4={^am2GOa=X)E^g`{HhAt%f-QiKwY(0N1 zzZ$)fsVR>RzLx!5&mk-2QP4D6%JAkeqU(vTrA&?lH=xHPrnw|E6to2$Y}ZeR(HBFE zHW{N!0iqD|Fd3K$np5*9*$jLTClf=`+CLU79Fu!JLM$zNAs>PyW0OTCDyj@JSpd}y zD979EhJWBPr16E|(e_Ep7aSbvC@QH%G9$KuYcB1 zev(~y2}iQ#{VQ-S+YU%je(FMYYxry3w>~ybPGA2G*?lx;=jh&5$w$0IU${@ntCmlK znYjYjOg%J)B2*E4GiJKkHZzS$1zZN9^xqterXk6e#m0TH2A(LK^6 z1JC!!cqI$EOj8vyZoS(lzzMMu7O@+wf4YE4al@^rmw$GmqVY>x`*c6$+ zLu^LgRI;>ahqQfRjrSA6_M6LGTi83}RI;TClg@~$JcStiLJhTo`wyxDNGxP-jfb&J|(wG&&9kZkpI z>JK?~!wn*7y-N{*8;E-fsN$9Mz8i4YDYJg8+$T#y_Nvm5nU4sD8>+m zCYtp7DNFQ;kq+)_V@3IAMUI;LkZYr{Z>)> zY;{E#3)QU&xXe>=n{sJVuWOVxi$t0iwa5s2TYfFH=o%s~Ymx&N{-xq_IM{{{rmpyX z!CcHb<{u8a;-ECN(r4aPuilW{LXIVJk%?TqN@rnL&k=4~vCFuXLxGxHu9ME!oiNYB zJ_){6OSd&6148+iEMgYPb$xJ28G19IxD{!OZ)*qpV$bl1v9K!o%)%MnHLYE?9C2{c zQT288yF2Riw^d!-qsvi@IUh{b4lC_O*u(D{W_FlE#Iq5~*qMIMKImIotk|4eUTr>A zcy0$4Nh1H?s52Z+H|e1oP$EO6iWD+&42iwgyS=OY;LW=+<#*$go26UfPbAX|_*+7|oiZ=(w6>OcAjE;{SE;_o z>k92suhoS+!nk!$Y>qAJ^!<~s?!cba6@LPllMd*xjo6y}*|t?`t#7E4LMo@tQwp$s_+$YoNhc6&;G8d5sTT z#{o3@ov5%g>A(gavpMenXZ6**!@BF)w9U4g*}LfzHZXhAZ_w7kKN5iJY9^i@&|wvb z21X~(HuYQ!Zgy{f`=ct)oozE^Mr=O%#pME4fHOZ}f}9Wo*!CbjC~&j{(>7C5xy4iX zDy^(`m5|!0ukdgXcp`N_}n3BpD! z%~ckvmwi3^d|7hLd=9VhO`2F-p_n=9tFf(3ArkzDcjnnpCNPohoo2iB{aWF zJa+S=2?4~mm*{3;?6!;Oc6NL=s$KQ1nX~dOEA8Q%z4y zOI#0o%PTH=YDd=}u6-lAAs&l;BDzgRchv4q@I@LoCQbN@WMbG%TK=ARtm60EhrUO) z*4$yai36j(KasHTm<#(R1HeiUVCoGt{{8*B{$1AQuj7fcwCCRgNVFQ=j;~0uMjwci z6yQ!eU)0yf+1ht|cNc$i*Sibk2&p{0T|NJWAu-S?4z!TkfB*gcx&B?%)@M5Q`u6p8 zlK54N(Nu8RVgQpUnj;jYp{&6t7BS_bWvVpaM5(ojLYIuQ)V5;Pd;T}Y zx8tbPcl-p&uTZg1=NGLz^@AjXam))QJ{Nc8Jo+8upFDo_ckJ?jXd2}LW8bPr8^m|l zPHp8m-1g<`Z0G9}@G>;H@4f-1r8;_xGt_2YgFfSA#i^}(&!?on`t&`;`rvG}6ZGs^ z0Nk9^PcHBtmPR{H>oA8dm`k23*}`02=|jzpMTa9i#@) zeG)jbM=pFz2udeXq$u+KxoPvM>5}?d`m+qG(Wg^KtdA&KhHEexq$ro?W zOjn-=M;k9thMqltnF0agRMyJexkQUQV`7j+$>%MKAsk zEJfeYDl+LU8fvujkAc@N!6#gZ=adlo@b>;$#rKc4rStoFU-AmYh_N>Xu1rVinqy6x za{e4nUS(uyy*BcSt#9k?@p$*pN44>C?fVyDiR98pQ0gBj@KL{^MvtK^zHO@wDpU@USdq@r*i}VNPS$z!z zmoCTQwOB3^U&rC?jK(wkTp+%8tu+fq7<}tnUs>uwz-qIPB_@Y=DCq?fKaz<9Oy&hi z=0cL;%=7Q-PNX_Edjxjnn?_)91qls0SB-v82}zk)HsLv4Q6Zj~G0`DI9|U z$vA2rP{;;r08vV7aIeY*5Nz5~O5cNt-yh8zAoUe)|2s^&7Rv^_TCFIUKb2+n4|Kh~ z4wCKbc#c^GCphI9P92=;^Xh1JX3t))&)*qf!(yg){74?pljpB{8LjJZbLpd{^fAXH zOu4)UBc`GZ0RuiUp)(5V=IpDb2F$X@P{)_?lpy|I-WDc@#s2$HD{YRLA7o2 zOQLGNs(18~o>r&-o}MsrGXIzSL_^8cj9yYio5zHW&D6-4jnj~o(S(D6jf2gUgOi1s z#hBBKf!&aelZo$tuh6AevbT4E`j5XwiC)sq%%0#s<`pG+WlJwp0!B_2dUb6^0%igx z-Tx__Vfp{GF|5mLaw>`_0S7Ikh!}}PM&c!v#B9S+5fQ0H5Nsed0R{tM1gcRQ;9+G! z5m7N1c|geaKDa_MFzyUzbBCu>lm4%qf6ek~iJ18BkRHa#|3&D&qxS7{_z5e1{iug% zE`&U(2( z(7Vl(8vbgTZVeMEWT=1WKeg>>?4*K#rN!60jN%UQ_#8Yuqtm^>IPGEJnPa*YJjEZ3 z@?7=B5)I=Hl-PvBhkC{*XD(Vf1h6};;o^WEt*H{dT%v=e&}d8haJ27&X*QumHT zS4e+PunJ314SN{3r}=7y%=d?!Hl520?B^Xbp3%+zK|$`2)SCmlEBN0(W?pkt{fB`3 z!{nN-(jPppH!{voRyKRcopVeq=crlEasD4t)*n^(x!d!t{PB*7eafruIXzGN&=);F;H&wqKar*r(m$*8v7Mf>+0H=xis&=3NCubx gp}R0sUvWudQAtHnY8scJv9Y0%374v>tG^o;01v8nc>n+a literal 0 HcmV?d00001 diff --git a/Aidaily_ca_tti_final_manuscript.tex b/Aidaily_ca_tti_final_manuscript.tex new file mode 100644 index 0000000..6db4b14 --- /dev/null +++ b/Aidaily_ca_tti_final_manuscript.tex @@ -0,0 +1,291 @@ +\documentclass[11pt]{article} +\usepackage[a4paper,margin=1in]{geometry} +\usepackage{fontspec} +\setmainfont{Times New Roman} +\usepackage{hyperref} +\hypersetup{colorlinks=true,linkcolor=blue,urlcolor=blue,citecolor=blue} +\usepackage{tabularx,array,booktabs} +\usepackage{enumitem} +\usepackage{fancyvrb} +\usepackage{titlesec} +\usepackage{setspace} +\usepackage{fancyhdr} +\pagestyle{fancy} +\fancyhf{} +\lhead{Transparency Drift in Human-AI Software Teams} +\rhead{\thepage} +\setstretch{1.08} +\setlength{\parskip}{0.6em} +\setlength{\parindent}{0pt} +\newcolumntype{Y}{>{\raggedright\arraybackslash}X} +\title{Transparency Drift in Human-AI Software Teams: A Confidence-Aware Team Transparency Index} +\date{June 26, 2026} +\begin{document} +\maketitle +\tableofcontents +\newpage +\section{Material Passport} +\begin{itemize}[leftmargin=*] +\item Origin Skill: academic-pipeline +\item Pipeline Entry Point: Stage 5 FINALIZE +\item Origin Date: 2026-06-26 +\item Prior Draft: \texttt{Aidaily\_ca\_tti\_manuscript\_draft.md} +\item Review Package: \texttt{Aidaily\_ca\_tti\_stage3\_review\_package.md} +\item Verification Status: FINALIZED / STAGE 4.5 FINAL INTEGRITY PASSED +\item Document Label: CA-TTI final manuscript +\item Contribution Type: Conceptual measurement framework with synthetic stress-test evidence +\item Source Materials: \texttt{Aidaily\_final\_manuscript.md}, \texttt{.context/ca\_tti\_synthetic\_validation\_plan.md}, \texttt{.context/ca\_tti\_synthetic\_output\_review.md}, San Diego synthetic experiment output, and Stage 3 review roadmap. +\end{itemize} +\section{Abstract} +\textbf{Background:} Software teams increasingly coordinate work through human communication, project-management artifacts, code repositories, code reviews, and AI-generated updates. Coding agents and AI assistants can generate pull requests, issue comments, documentation, summaries, and decision records. These artifacts can make a project appear more complete and traceable even when human review, shared understanding, and accountability are weakening. +\textbf{Objective:} This paper introduces CA-TTI, a confidence-aware Team Transparency Index for detecting transparency drift in human-AI software teams. Transparency drift is defined as a growing divergence between visible project artifacts and the human team's shared understanding, review depth, and ability to explain or control the work being produced. +\textbf{Methods:} We revise an earlier Team Transparency Index based on coverage, consistency, consensus, timeliness, and completeness. CA-TTI preserves these artifact-oriented components but changes the output from a single score into a multi-signal framework: artifact score, confidence, trend state, and Human-Agent Alignment Gap (HAG). HAG is treated as a family of alignment indicators, not as a single validated latent variable. We evaluate initial measurement behavior using a synthetic stress test across five controlled scenarios. +\textbf{Results:} In the synthetic prototype, raw TTI missed all warning cases in artifact drift and fluent hallucination scenarios. CA-TTI warned in all such trials, with mean warning lead times of 3.0 steps for artifact drift and 1.5 steps for fluent hallucination. In a noisy-interaction scenario without true artifact failure, raw TTI warned in 7 of 8 trials, while CA-TTI produced no warnings. Ablation checks indicated that removing trend logic substantially reduced warning lead time in the two failure scenarios. +\textbf{Conclusion:} CA-TTI is proposed as an early-warning measurement framework, not as a validated productivity metric. Its central claim is that transparency in human-AI software teams must be interpreted as a multi-signal condition. Artifact completeness can improve while shared human understanding declines. Future work should implement event-level datasets, test inter-rater reliability, and validate the framework in real teams. +\textbf{Keywords:} human-AI software teams; team transparency; AI agents; software traceability; Agile software development; shared mental models; measurement framework; transparency drift +\section{1. Introduction} +Software teams do not coordinate only through formal records. They rely on stand-up meetings, chat threads, issue trackers, pull requests, code reviews, release notes, documentation, and informal memory. These sources create a distributed picture of work. A decision may begin in a meeting, be clarified in chat, appear indirectly in a pull request, and never be reflected in the issue tracker. This fragmentation creates a persistent gap between what the team informally knows and what the project system formally records. +Earlier versions of the Team Transparency Index (TTI) were designed to measure whether communication, project artifacts, and confirmed team knowledge were aligned. The original use case was an AI mediator for Agile teams: a conversational system would ingest stand-ups, chat, issue data, and Git metadata; extract candidate decisions and action items; detect mismatches; and request role-aware confirmation before writing back to the project record. In that framing, TTI was mainly an outcome measure for evaluating whether the mediator improved transparency. +The rise of coding agents changes the measurement problem. AI systems no longer only summarize or remind. They can generate code, open pull requests, update issues, draft documentation, propose decisions, and produce fluent explanations. Evidence from AI pair-programming and agentic code-review research suggests that AI systems are becoming part of everyday software workflows, although their contributions, adoption patterns, and review needs differ from human work (Peng et al., 2023; Zhong et al., 2026). This can improve visible traceability, but it can also create a new kind of transparency failure. The artifact layer may become more complete while the team loses human understanding, review depth, accountability, or trust. +This paper calls that failure mode \textbf{transparency drift}: a gradual divergence between visible project artifacts and the shared understanding of the human team. Transparency drift matters because artifact fluency can look like coordination. A project may show more comments, cleaner issue records, faster summaries, and more complete decision logs while team members are less able to explain why decisions were made, who endorsed them, whether AI-generated changes were deeply reviewed, or how errors should be corrected. +The central contribution of this paper is CA-TTI, a revised measurement framework for human-AI software teams. The framework is not a validated field index yet. It is a conceptual and methodological proposal supported by synthetic stress-test evidence. The key shift is: +\begin{Verbatim}[fontsize=\small] +CA-TTI = artifact score + confidence + trend state + Human-Agent Alignment Gap +\end{Verbatim} +CA-TTI is not intended to replace human judgment or rank individuals. It is an early-warning framework for team-level inquiry. Its purpose is to show when an apparently orderly system may be drifting away from shared human understanding. +\section{2. Related Work and Positioning} +CA-TTI sits between four research streams: Agile coordination, software traceability, human-AI teaming, and governance of workplace monitoring. +Agile software development depends on frequent communication and shared context. Daily stand-up meetings can support awareness, coordination, and monitoring, but prior work also shows that their value depends on team context and meeting quality (Stray et al., 2016, 2017, 2020). Agile transparency therefore cannot be inferred from the presence of ceremonies alone. +Traceability research raises a parallel issue. Software traceability is valuable, but it is often created ad hoc and after the fact, which limits its practical benefit (Cleland-Huang et al., 2014). Socio-technical congruence research similarly links coordination needs to actual coordination patterns in software teams (Cataldo et al., 2008). These literatures show that software transparency is not only a documentation problem. It is a relationship between work dependencies, communication, records, and shared understanding. +Code review adds a further signal. Code review is not merely a gate for defect detection. It also supports knowledge transfer, maintainability, and shared standards; developers judge review quality through factors such as feedback usefulness, clarity, and reviewer expertise (Kononenko et al., 2016). When AI agents participate in code review, the question is not only whether feedback is syntactically correct, but whether humans still understand, contest, and integrate that feedback in meaningful ways. +Human-AI teaming research provides the theoretical bridge. Human-AI teams require more than a tool-user relationship; they require coordination around shared goals, roles, communication, trust, and team cognition (Berretta et al., 2023). Shared mental models are especially relevant because a team can coordinate effectively only when members maintain compatible expectations about tasks, roles, and system behavior (Andrews et al., 2023). Empirical work on human-agent teams also suggests that communication, explicitly shared goals, trust, and perceived team cognition shape performance and collaboration (Schelble et al., 2022). +AI-supported project work adds another layer. Recent work on requirements engineering, cognitive agents, and Agile project-management support suggests that AI and machine-learning systems can extract structured requirements information and simulate or support Agile project-management roles (Cinkusz et al., 2025; Umar et al., 2025). Human-centered AI research warns that useful automation must preserve human control, safety, and trust (Shneiderman, 2020). Workplace surveillance research adds a governance concern: measurement systems can become instruments of monitoring and pressure if they expose individual behavior without appropriate safeguards (Ball, 2021). +These strands suggest a measurement problem rather than only a tool-building problem. The question is not simply whether AI can produce more complete project artifacts. The harder question is whether a team remains genuinely transparent when AI participates in producing those artifacts. +\section{3. Construct Boundary: What CA-TTI Measures} +This paper uses transparency in a team-level, socio-technical sense. It does not equate transparency with explainability of an AI model or with the number of project artifacts. Four forms of transparency are relevant: +\begin{center}\small +\begin{tabularx}{\linewidth}{Y|Y|Y} +\toprule +Transparency form & Question & Example signal \\ +\midrule +Artifact transparency & Are project records linked, current, and complete? & Issue, pull request, commit, and decision-record links \\ +Process transparency & Is it clear how work moved from discussion to decision to implementation? & Trace from meeting or chat decision to implementation artifact \\ +Epistemic transparency & Can affected humans explain and challenge the work? & Human review depth, confirmation, correction, and shared explanation \\ +Governance transparency & Are provenance, access, consent, and accountability clear? & AI provenance labels, appeal paths, access rules \\ +\bottomrule +\end{tabularx} +\end{center} +CA-TTI is designed for the intersection of these forms. The original artifact-oriented TTI mostly measured artifact transparency. CA-TTI keeps that core but adds confidence, trend, and HAG so that artifact improvement is not mistaken for full team transparency. +The intended use is a team-level diagnostic and audit protocol. It can feed a dashboard, but the dashboard is not the primary contribution. The primary contribution is an interpretation rule: CA-TTI warnings should trigger team inquiry, not sanctions or individual performance assessment. A team uses the index to ask: "Do our artifacts still reflect what we jointly understand, review, and control?" +\section{4. From the original artifact-oriented TTI to CA-TTI} +\subsection{4.1 Original Artifact-Oriented TTI: Artifact Transparency} +The original TTI used five components: +\begin{Verbatim}[fontsize=\small] +TTI = 0.25*COV + 0.25*CON + 0.20*CSN + 0.15*TML + 0.15*CMP +\end{Verbatim} +\begin{center}\small +\begin{tabularx}{\linewidth}{Y|Y} +\toprule +Component & Meaning \\ +\midrule +COV & Coverage: eligible communication-mentioned tasks or decisions linked to an issue, pull request, Git artifact, or decision record. \\ +CON & Consistency: status, ownership, blocker, and decision claims match structured records or are explicitly reconciled. \\ +CSN & Consensus: high-impact decisions meet predefined role-confirmation thresholds before writeback. \\ +TML & Timeliness: documented updates occur soon enough after the relevant event. \\ +CMP & Completeness: action items and decisions include required metadata such as who, what, and when. \\ +\bottomrule +\end{tabularx} +\end{center} +This score is useful because it makes artifact transparency operational. However, it remains vulnerable to two problems. First, the score can become overconfident under sparse or uneven data. If only one source is available, a clean-looking score may reflect missing evidence rather than real alignment. Second, the score can reward documentation hygiene even when human alignment is weakening. AI agents may improve links, summaries, and completeness while reducing review depth or obscuring responsibility. +\subsection{4.2 CA-TTI: Multi-Signal Transparency} +CA-TTI keeps the five artifact components but changes the output format. Instead of returning a single score, it returns four signals: +\begin{Verbatim}[fontsize=\small] +artifact_score +confidence +trend_state +human_agent_alignment_gap +\end{Verbatim} +The artifact score answers whether visible project records are linked, consistent, confirmed, timely, and complete. Confidence answers how much evidence supports that score. Trend state answers whether the system is improving, stable, deteriorating, noisy, or insufficiently observed. The Human-Agent Alignment Gap answers whether AI-generated artifacts remain aligned with human understanding, review, endorsement, and control. +This separation matters because the same artifact score can have different meanings. A high artifact score with high confidence and low HAG is a strong signal. A high artifact score with low confidence is fragile. A high artifact score with rising HAG is the core risk case: artifacts look clean while shared understanding deteriorates. +\subsection{4.3 Signal Flow} +\begin{Verbatim}[fontsize=\small] +team events + -> communication, issues, pull requests, reviews, decisions, AI actions + -> artifact score / confidence / trend / HAG + -> warning state + -> team inquiry and governance response +\end{Verbatim} +CA-TTI therefore should not be read as a scalar ranking. It is a bundle of diagnostic signals. The warning state is useful only if the team investigates the underlying components. +\section{5. Human-Agent Alignment Gap} +The Human-Agent Alignment Gap (HAG) is the distance between what agents produce or record and what the human team has actually reviewed, endorsed, understood, or accepted. +HAG is not an anti-AI measure. It does not treat agent participation as harmful by default. It asks whether AI contribution is growing faster than the team's capacity for review, confirmation, and shared understanding. +HAG should be treated as a family of subdimensions: +\begin{center}\small +\begin{tabularx}{\linewidth}{Y|Y|Y|Y} +\toprule +HAG subdimension & Definition & Observable indicator & Current prototype coverage \\ +\midrule +Confirmation debt & AI-created or AI-modified claims lack explicit human confirmation. & Unconfirmed decision records, auto-updated tickets, unendorsed summaries & Partial \\ +Review-depth gap & AI-generated work receives shallower review than comparable human work. & Review rounds, comment depth, test discussion, reviewer expertise & Not yet implemented \\ +Attribution ambiguity & The team cannot tell whether a claim came from a human, an agent, or a mixed process. & Missing provenance label, unclear author chain & Not yet implemented \\ +Explanation gap & Affected humans cannot explain why a decision or change was made. & Post-hoc explanation checks, reviewer challenge outcomes & Not yet implemented \\ +Correction gap & AI-generated artifacts require more correction, revert, or clarification. & Reverts, follow-up corrections, rejected suggestions & Partial \\ +Trust and safety divergence & Artifacts improve while trust, psychological safety, or willingness to challenge declines. & Team survey or retrospective signal & Not yet implemented \\ +\bottomrule +\end{tabularx} +\end{center} +The current synthetic prototype does not implement full HAG. It implements a narrower \texttt{hag\_proxy}, a hallucination-alignment proxy derived from unsupported claims, confidence-artifact mismatch, and missing evidence. This is a useful first stress-test component, but it should not be interpreted as a complete operationalization of human-agent alignment. +\section{6. Synthetic Stress-Test Design} +The present draft uses a synthetic stress test to evaluate whether CA-TTI behaves sensibly under controlled failure modes. This is not a field validation study. It is a measurement-behavior test. +The prototype compared a raw TTI-like score against a CA-TTI score. The generator produced 8 trials and 10 steps for each of five scenarios, yielding 400 trajectory-level observations. Each row included: +\begin{Verbatim}[fontsize=\small] +scenario +trial_id +step +raw_tti +artifact_score +confidence +unsupported_claim_rate +evidence_coverage +actual_quality +failure_label +\end{Verbatim} +The scenarios were: +\begin{center}\small +\begin{tabularx}{\linewidth}{Y|Y} +\toprule +Scenario & Purpose \\ +\midrule +Clean baseline & Raw TTI and artifact-centered signals should agree. \\ +Artifact drift & Artifact quality deteriorates before interaction quality visibly fails. \\ +Fluent hallucination & Output remains fluent and confident while artifact quality and evidence coverage collapse. \\ +Low-confidence good artifacts & Useful artifacts exist but system confidence is muted. \\ +Noisy interaction with stable artifacts & Interaction-level noise should not be treated as artifact failure. \\ +\bottomrule +\end{tabularx} +\end{center} +The prototype scorer computed trend from a four-step artifact-history window. A negative artifact slope reduced the trend score; fewer than four observations defaulted to a neutral trend score. The prototype \texttt{hag\_proxy} combined unsupported-claim rate, confidence-artifact mismatch, and evidence-coverage gap: +\begin{Verbatim}[fontsize=\small] +hag_proxy = + 0.52 * unsupported_claim_rate ++ 0.30 * max(0, confidence - artifact_score) ++ 0.18 * max(0, 1 - evidence_coverage) +\end{Verbatim} +Confidence was calibrated downward when \texttt{hag\_proxy} rose: +\begin{Verbatim}[fontsize=\small] +calibrated_confidence = confidence * (1 - 0.65 * hag_proxy) +\end{Verbatim} +The prototype CA-TTI score was: +\begin{Verbatim}[fontsize=\small] +ca_tti_score = + 0.45 * artifact_score ++ 0.20 * calibrated_confidence ++ 0.20 * trend_score ++ 0.15 * (1 - hag_proxy) +\end{Verbatim} +Warnings were generated when: +\begin{Verbatim}[fontsize=\small] +ca_tti_score < 0.58 +or hag_proxy >= 0.42 +or trend_score <= 0.42 +\end{Verbatim} +The raw baseline used the same numeric warning threshold for the raw TTI-like score: +\begin{Verbatim}[fontsize=\small] +raw_tti < 0.58 +\end{Verbatim} +These thresholds were chosen as prototype stress-test settings, not as validated field cutoffs. A deployed system would require calibration by domain, team workflow, and evidence availability. +\section{7. Synthetic Results} +The results are summarized below. +\begin{center}\small +\begin{tabularx}{\linewidth}{Y|Y|Y|Y|Y} +\toprule +Scenario & Failure Trials & Raw Warn Trials & CA-TTI Warn Trials & Mean CA-TTI Lead \\ +\midrule +Artifact drift & 8/8 & 0/8 & 8/8 & 3.0 steps \\ +Fluent hallucination & 8/8 & 0/8 & 8/8 & 1.5 steps \\ +Clean baseline & 0/8 & 0/8 & 0/8 & n/a \\ +Low-confidence good artifacts & 0/8 & 1/8 & 0/8 & n/a \\ +Noisy interaction with stable artifacts & 0/8 & 7/8 & 0/8 & n/a \\ +\bottomrule +\end{tabularx} +\end{center} +In the two failure scenarios, raw TTI did not warn, while CA-TTI warned in every trial. In the noisy but stable scenario, raw TTI produced warnings in most trials, while CA-TTI produced none. In the low-confidence good-artifact scenario, CA-TTI did not treat low confidence alone as failure when artifacts and evidence remained strong. +An ablation check using the same scored observations suggests that trend logic contributed substantially to early warning. Removing HAG did not change warning counts in this small prototype, but removing trend reduced lead time below the failure step in both failure scenarios. The check is preliminary because it was not pre-registered and uses the same synthetic data. +\begin{center}\small +\begin{tabularx}{\linewidth}{Y|Y|Y|Y|Y} +\toprule +Failure scenario & Full CA-TTI lead & No HAG lead & No trend lead & Artifact-only lead \\ +\midrule +Artifact drift & 3.0 & 3.0 & -0.25 & 1.0 \\ +Fluent hallucination & 1.5 & 1.5 & -0.88 & 0.25 \\ +\bottomrule +\end{tabularx} +\end{center} +These results support only a bounded claim: separating artifact score, confidence calibration, trend, and alignment-gap signals can make synthetic failure modes visible that a single raw score can miss. They do not establish construct validity, external validity, or real-world predictive accuracy. +\section{8. Use, Misuse, and Deployment Vignette} +CA-TTI should be used as a team-level diagnostic and early-warning framework. It should not be used to score individual developers, rank teams, or evaluate personal performance. Such use would distort behavior and create surveillance risk. +The governance rule is: +\begin{Verbatim}[fontsize=\small] +A team should not be considered more transparent if artifact transparency improves while psychological safety, trust, or human-agent alignment declines. +\end{Verbatim} +Practical deployment requires safeguards: +\begin{enumerate}[leftmargin=*] +\item Report CA-TTI at the team level, not the individual level. +\item Separate artifact score from confidence, trend, and HAG. +\item Preserve provenance for AI-created or AI-modified records. +\item Require explicit human confirmation for high-impact decisions. +\item Treat psychological safety and workload as hard governance constraints. +\item Allow participants to challenge, correct, or reverse AI-mediated records. +\item Avoid manager-facing dashboards that expose individual prompt behavior. +\item Retain raw event data only as long as needed for team-level audit. +\item Treat warnings as inquiry triggers, not sanctions. +\end{enumerate} +A safe deployment vignette illustrates the intended use. During a sprint review, a team sees that artifact score has risen from 0.71 to 0.84 because more pull requests, issue links, and decision records are present. Confidence is moderate, but HAG is rising because several agent-generated summaries were accepted without review and developers cannot explain the rationale behind two decisions. The correct response is not to identify a low-performing developer. The correct response is a team inquiry: review which AI-generated records require confirmation, add provenance labels, ask affected developers to explain or revise the decision records, and adjust review policy for future agent-generated changes. +This use case also shows how gaming should be handled. A team could try to increase confirmations without improving understanding. CA-TTI therefore should not reward confirmation count alone. Confirmation must be linked to review depth, provenance, correction rights, and the ability of affected humans to challenge the record. +\section{9. Discussion} +The shift from the original artifact-oriented TTI to CA-TTI changes the paper's contribution. The original contribution was a protocol for evaluating a conversational AI mediator. The revised contribution is a measurement framework for human-AI software teams. +This change makes the paper more durable. Tool designs will change quickly as AI agents evolve. A measurement problem will remain: how can teams know whether AI participation is improving shared transparency or merely improving the appearance of traceability? +CA-TTI addresses that problem by refusing to compress transparency into one score. A high artifact score with low confidence should not be interpreted like a high artifact score with rich evidence. A high artifact score with rising HAG should not be interpreted as success. A temporary noisy transition should not be treated like confirmed transparency decline. These distinctions are the main value of the framework. +The framework also clarifies how future validation should proceed. A field study should not only ask whether CA-TTI increases during an intervention. It should ask whether CA-TTI predicts cases where teams report lower alignment, lower trust, lower psychological safety, or weaker shared mental models despite improved artifact completeness. +\section{10. Limitations} +This paper has five major limitations. +First, the synthetic data are not real Agile data. They test measurement behavior under designed scenarios but cannot establish construct validity. +Second, the current prototype is trajectory-level rather than event-level. A fuller version should generate or collect event-level data with fields such as event type, source, actor type, linked artifact, confirmation status, delay, metadata completeness, review depth, autonomy level, correction events, trust signal, psychological safety signal, and missingness flag. +Third, HAG is only partially operationalized in the current prototype. The implemented \texttt{hag\_proxy} focuses on unsupported claims and confidence-evidence mismatch. The broader construct should include review depth, confirmation debt, attribution ambiguity, correction/revert rate, explanation gap, trust decline, and psychological safety decline. +Fourth, the thresholds used in this prototype are not validated field cutoffs. They are stress-test settings. Real deployment would require calibration, sensitivity analysis, and stakeholder review. +Fifth, the literature base is stronger than in the initial draft but still selective. A full journal submission should include a broader review of AI coding agents, team cognition measurement, software repository mining, and responsible workplace AI governance. +\section{11. Future Work} +Future work should proceed in three stages. +First, improve the synthetic generator so it produces event-level software-team records rather than only trajectories. This would make the metric easier to explain and closer to the eventual field setting. +Second, conduct coder-based plausibility review. Human reviewers should inspect generated scenarios and judge whether the synthetic events plausibly represent healthy teams, documentation-only improvement, human-agent drift, noisy transition, and true transparency decline. Inter-rater reliability should be reported. +Third, run a small field feasibility study. The goal should not be causal proof. The goal should be to test whether CA-TTI can be computed reliably, whether HAG can be coded consistently, and whether team members find the outputs meaningful rather than intrusive. +\section{12. Conclusion} +Human-AI software teams need a way to detect transparency drift before it becomes visible as project failure. Raw artifact scores are not enough because AI agents can improve the visible record while weakening shared human understanding. CA-TTI responds by separating artifact transparency, confidence, trend, and Human-Agent Alignment Gap. +The first synthetic stress test supports the plausibility of this framing. CA-TTI detected artifact drift and fluent hallucination scenarios that raw TTI missed, while avoiding false warnings in noisy but stable conditions. Ablation checks suggest that trend logic is important for early warning in the current prototype. These findings are preliminary, but they show why a single transparency score is not sufficient for human-AI software teams. +CA-TTI should therefore be understood as an early-warning measurement framework. Its purpose is not to declare that a team is productive or well-managed. Its purpose is to ask whether the team still understands, reviews, confirms, and controls the work that humans and agents are producing together. +\section{Declarations} +\subsection{Data Availability} +The synthetic stress-test outputs used in this manuscript are available in the accompanying project experiment output directory. The current manuscript reports only synthetic measurement-behavior results and does not include human-subject data. +\subsection{Funding} +No external funding is declared for this manuscript draft. +\subsection{Conflicts of Interest} +The author declares no conflicts of interest for this manuscript draft. +\subsection{AI Disclosure} +This paper was prepared with the assistance of AI-powered academic writing tools. The AI pipeline included research framing, structure planning, draft writing, reviewer simulation, revision planning, citation verification, integrity checking, and formatting support. All content, arguments, and conclusions were directed and reviewed by the author. The author takes full responsibility for the accuracy and integrity of this work. +\section{References} +Andrews, R. W., Lilly, J. M., Srivastava, D., \& Feigh, K. M. (2023). The role of shared mental models in human-AI teams: A theoretical review. \emph{Theoretical Issues in Ergonomics Science, 24}(2), 129-175. https://doi.org/10.1080/1463922X.2022.2061080 +Ball, K. (2021). \emph{Electronic monitoring and surveillance in the workplace: Literature review and policy recommendations}. Publications Office of the European Union. https://doi.org/10.2760/5137 +Berretta, S., Tausch, A., Ontrup, G., Gilles, B., Peifer, C., \& Kluge, A. (2023). Defining human-AI teaming the human-centered way: A scoping review and network analysis. \emph{Frontiers in Artificial Intelligence, 6}, Article 1250725. https://doi.org/10.3389/frai.2023.1250725 +Cataldo, M., Herbsleb, J. D., \& Carley, K. M. (2008). Socio-technical congruence. In \emph{Proceedings of the Second ACM-IEEE International Symposium on Empirical Software Engineering and Measurement} (pp. 2-11). ACM. https://doi.org/10.1145/1414004.1414008 +Cinkusz, K., Chudziak, J. A., \& Niewiadomska-Szynkiewicz, E. (2025). Cognitive agents powered by large language models for agile software project management. \emph{Electronics, 14}(1), Article 87. https://doi.org/10.3390/electronics14010087 +Cleland-Huang, J., Gotel, O. C. Z., Huffman Hayes, J., Mäder, P., \& Zisman, A. (2014). Software traceability: Trends and future directions. In \emph{Future of Software Engineering Proceedings} (pp. 55-69). ACM. https://doi.org/10.1145/2593882.2593891 +Kononenko, O., Baysal, O., \& Godfrey, M. W. (2016). Code review quality. In \emph{Proceedings of the 38th International Conference on Software Engineering} (pp. 1028-1038). ACM. https://doi.org/10.1145/2884781.2884840 +Peng, S., Kalliamvakou, E., Cihon, P., \& Demirer, M. (2023). \emph{The impact of AI on developer productivity: Evidence from GitHub Copilot}. arXiv. https://arxiv.org/abs/2302.06590 +Schelble, B. G., Flathmann, C., McNeese, N. J., Freeman, G., \& Mallick, R. (2022). Let's think together! Assessing shared mental models, performance, and trust in human-agent teams. \emph{Proceedings of the ACM on Human-Computer Interaction, 6}(GROUP), Article 13, 1-29. https://doi.org/10.1145/3492832 +Shneiderman, B. (2020). Human-centered artificial intelligence: Reliable, safe \& trustworthy. \emph{International Journal of Human-Computer Interaction, 36}(6), 495-504. https://doi.org/10.1080/10447318.2020.1741118 +Stray, V., Moe, N. B., \& Bergersen, G. R. (2017). Are daily stand-up meetings valuable? A survey of developers in software teams. In H. Baumeister, H. Lichter, \& M. Riebisch (Eds.), \emph{Agile Processes in Software Engineering and Extreme Programming} (pp. 274-281). Springer. https://doi.org/10.1007/978-3-319-57633-6\_20 +Stray, V., Moe, N. B., \& Sjoberg, D. I. K. (2020). Daily stand-up meetings: Start breaking the rules. \emph{IEEE Software, 37}(3), 70-77. https://doi.org/10.1109/MS.2018.2875988 +Stray, V., Sjøberg, D. I. K., \& Dybå, T. (2016). The daily stand-up meeting: A grounded theory study. \emph{Journal of Systems and Software, 114}, 101-124. https://doi.org/10.1016/j.jss.2016.01.004 +Umar, M. A., Lano, K., \& Abubakar, A. K. (2025). Automated requirements engineering framework for agile model-driven development. \emph{Frontiers in Computer Science, 7}, Article 1537100. https://doi.org/10.3389/fcomp.2025.1537100 +Zhong, S., Noei, S., Zou, Y., \& Adams, B. (2026). \emph{Human-AI synergy in agentic code review}. arXiv. https://arxiv.org/abs/2603.15911 +\end{document} \ No newline at end of file diff --git a/Aidaily_ca_tti_manuscript_draft.md b/Aidaily_ca_tti_manuscript_draft.md new file mode 100644 index 0000000..347a390 --- /dev/null +++ b/Aidaily_ca_tti_manuscript_draft.md @@ -0,0 +1,254 @@ +# Transparency Drift in Human-AI Software Teams: A Confidence-Aware Team Transparency Index + +## Material Passport + +- Origin Skill: academic-pipeline +- Pipeline Entry Point: Stage 2 WRITE +- Origin Date: 2026-06-26 +- Verification Status: STAGE 2.5 PRE-REVIEW AUDITED / CORRECTIONS APPLIED / FIELD VALIDITY PENDING +- Document Label: CA-TTI draft manuscript +- Source Materials: `Aidaily_final_manuscript.md`, `.context/ca_tti_synthetic_validation_plan.md`, `.context/ca_tti_synthetic_output_review.md`, and San Diego synthetic experiment output. + +## Abstract + +**Background:** Software teams increasingly coordinate work through a mixture of human communication, project-management artifacts, code repositories, and AI-generated updates. AI agents can produce pull requests, summaries, issue comments, decision records, and status updates that make the visible artifact layer appear more complete. However, artifact completeness does not necessarily mean that a team has preserved shared understanding, human review, trust, or psychological safety. A team may look more transparent while becoming less aligned. + +**Objective:** This paper introduces CA-TTI, a confidence-aware Team Transparency Index for detecting transparency drift in human-AI software teams. Rather than treating transparency as a single score, CA-TTI separates artifact transparency, confidence, temporal trend, and the Human-Agent Alignment Gap (HAG). + +**Methods:** We extend an earlier Team Transparency Index based on coverage, consistency, consensus, timeliness, and completeness. CA-TTI preserves these five artifact-oriented components but adds confidence calibration, trend-sensitive warning logic, and HAG as a separate signal for cases where AI-generated artifacts outpace human confirmation, review, or shared understanding. We evaluate the measurement behavior using synthetic stress tests that compare a raw TTI-like score against CA-TTI across scenarios including clean baseline behavior, artifact drift, fluent hallucination, low-confidence but good artifacts, and noisy interaction with stable artifacts. + +**Results:** In the synthetic prototype, raw TTI missed all warning cases in artifact drift and fluent hallucination scenarios. CA-TTI warned in all such trials, with mean warning lead times of 3.0 steps for artifact drift and 1.5 steps for fluent hallucination. In a noisy-interaction scenario without true artifact failure, raw TTI produced warnings in 7 of 8 trials, while CA-TTI produced no warnings. These results do not establish real-world validity, but they show that separating score, confidence, trend, and human-agent alignment can expose failure modes that a single score can miss. + +**Conclusion:** CA-TTI is proposed as an early-warning measurement framework, not as a universal productivity score. Its central claim is that transparency in human-AI software teams must be interpreted as a multi-signal condition: visible artifacts may improve while shared human understanding declines. Future work should implement event-level datasets, test inter-rater reliability, and validate the index in real software teams. + +**Keywords:** human-AI software teams; team transparency; AI agents; software traceability; Agile software development; measurement model; synthetic validation; transparency drift + +## 1. Introduction + +Software teams do not coordinate only through formal records. They rely on stand-up meetings, chat threads, issue trackers, pull requests, code reviews, release notes, and informal memory. These sources create a distributed picture of work. A decision may begin in a meeting, be clarified in chat, appear indirectly in a pull request, and never be reflected in the issue tracker. This fragmentation creates a persistent gap between what the team informally knows and what the project system formally records. + +Earlier versions of the Team Transparency Index (TTI) were designed to measure whether communication, project artifacts, and confirmed team knowledge were aligned. The original use case was an AI mediator for Agile teams: a conversational system would ingest stand-ups, chat, issue data, and Git metadata; extract candidate decisions and action items; detect mismatches; and request role-aware confirmation before writing back to the project record. In that framing, TTI was mainly an outcome measure for evaluating whether the mediator improved transparency. + +The rise of coding agents changes the measurement problem. AI systems no longer only summarize or remind. They can generate code, open pull requests, update issues, draft documentation, propose decisions, and produce fluent explanations. This can improve visible traceability, but it can also create a new kind of transparency failure. The artifact layer may become more complete while the team loses human understanding, review depth, accountability, or trust. + +This paper calls that failure mode **transparency drift**: a gradual divergence between visible project artifacts and the shared understanding of the human team. Transparency drift matters because artifact fluency can look like coordination. A project may show more comments, cleaner issue records, faster summaries, and more complete decision logs while team members are less able to explain why decisions were made, who endorsed them, whether AI-generated changes were deeply reviewed, or how errors should be corrected. + +The central contribution of this paper is CA-TTI, a revised measurement framework for human-AI software teams. The key shift is simple: + +```text +CA-TTI = artifact score + confidence + trend + Human-Agent Alignment Gap +``` + +CA-TTI is not intended to replace human judgment or rank individuals. It is an early-warning framework for team-level transparency. Its purpose is to show when an apparently orderly system may be drifting away from shared human understanding. + +## 2. Background and Motivation + +Agile software development depends on frequent communication and shared context. Daily stand-up meetings can support awareness, coordination, and monitoring, but prior work also shows that their value depends on team context and meeting quality (Stray et al., 2016, 2017, 2020). Agile transparency therefore cannot be inferred from the presence of ceremonies alone. + +Traceability research raises a parallel issue. Software traceability is valuable, but it is often created ad hoc and after the fact, which limits its practical benefit (Cleland-Huang et al., 2014). A team can have many artifacts and still lack a reliable chain from discussion to decision to implementation. Conversely, informal understanding may exist but remain invisible to the project record. + +AI-supported project work adds another layer. Recent work on requirements engineering, AI agents, and Agile project-management support suggests that AI and machine-learning systems can extract structured requirements information and simulate or support Agile project-management roles (Cinkusz et al., 2025; Umar et al., 2025). Human-centered AI research also warns that useful automation must preserve human control, safety, and trust (Shneiderman, 2020). Workplace surveillance research adds a governance concern: measurement systems can become instruments of monitoring and pressure if they expose individual behavior without appropriate safeguards (Ball, 2021). + +These strands suggest a measurement problem rather than only a tool-building problem. The question is not simply whether AI can produce more complete project artifacts. The harder question is whether a team remains genuinely transparent when AI participates in producing those artifacts. + +## 3. From the original artifact-oriented TTI to CA-TTI + +### 3.1 Original Artifact-Oriented TTI: Artifact Transparency + +The original TTI used five components: + +```text +TTI = 0.25*COV + 0.25*CON + 0.20*CSN + 0.15*TML + 0.15*CMP +``` + +| Component | Meaning | +| --- | --- | +| COV | Coverage: eligible communication-mentioned tasks or decisions linked to an issue, pull request, Git artifact, or decision record. | +| CON | Consistency: status, ownership, blocker, and decision claims match structured records or are explicitly reconciled. | +| CSN | Consensus: high-impact decisions meet predefined role-confirmation thresholds before writeback. | +| TML | Timeliness: documented updates occur soon enough after the relevant event. | +| CMP | Completeness: action items and decisions include required metadata such as who, what, and when. | + +This score is useful because it makes artifact transparency operational. However, it remains vulnerable to two problems. + +First, the score can become overconfident under sparse or uneven data. If only one source is available, a clean-looking score may reflect missing evidence rather than real alignment. + +Second, the score can reward documentation hygiene even when human alignment is weakening. AI agents may improve links, summaries, and completeness while reducing review depth or obscuring responsibility. + +### 3.2 CA-TTI: Multi-Signal Transparency + +CA-TTI keeps the five artifact components but changes the output format. Instead of returning a single score, it returns four signals: + +```text +artifact_score +confidence +trend_state +human_agent_alignment_gap +``` + +The artifact score answers: Are the visible project records linked, consistent, confirmed, timely, and complete? + +Confidence answers: How much evidence supports the score? Confidence should fall when event counts are low, source coverage is partial, consent gaps are large, coder reliability is unknown, or tool streams are missing. + +Trend state answers: Is the system improving, stable, deteriorating, noisy, or insufficiently observed? Trend matters because transparency problems often appear gradually. A single sprint can be noisy; repeated deterioration is more informative. + +The Human-Agent Alignment Gap answers: Are AI-generated artifacts still aligned with human understanding, review, and control? + +## 4. Human-Agent Alignment Gap + +The Human-Agent Alignment Gap (HAG) is the distance between what agents produce or record and what the human team has actually reviewed, endorsed, understood, or accepted. + +HAG is not an anti-AI measure. It does not treat agent participation as harmful by default. It asks whether AI contribution is growing faster than the team's capacity for review, confirmation, and shared understanding. + +Examples of high-HAG situations include: + +- AI-generated updates are written without explicit human confirmation. +- Pull requests created by agents are merged after shallow review. +- Decision records are produced by an agent but cannot be explained by the affected team members. +- Correction or revert rates increase on AI-generated artifacts. +- Attribution becomes unclear: the team cannot tell whether a claim came from a human, an agent, or a mixed process. +- Documentation improves while trust or psychological safety declines. + +HAG should initially remain separate from the artifact score. Combining it too early into one final number would hide the most important interpretive distinction. A team can have high artifact transparency and high HAG at the same time. That combination is exactly the risk case this paper aims to detect. + +## 5. Synthetic Stress-Test Design + +The present draft uses a synthetic stress test to evaluate whether CA-TTI behaves sensibly under controlled failure modes. This is not a field validation study. It is a measurement behavior test. + +The prototype compared a raw TTI-like score against a CA-TTI score. The synthetic generator produced trajectories for five scenarios: + +| Scenario | Purpose | +| --- | --- | +| Clean baseline | Raw TTI and artifact-centered signals should agree. | +| Artifact drift | Artifact quality deteriorates before interaction quality visibly fails. | +| Fluent hallucination | Output remains fluent and confident while artifact quality and evidence coverage collapse. | +| Low-confidence good artifacts | Useful artifacts exist but system confidence is muted. | +| Noisy interaction with stable artifacts | Interaction-level noise should not be treated as artifact failure. | + +The implemented prototype used trajectory-level observations rather than full event-level Agile records. Each row included: + +```text +raw_tti +artifact_score +confidence +unsupported_claim_rate +evidence_coverage +actual_quality +failure_label +``` + +The prototype scorer computed: + +```text +ca_tti_score = + 0.45 * artifact_score ++ 0.20 * calibrated_confidence ++ 0.20 * trend_score ++ 0.15 * (1 - hag) +``` + +Warnings were generated when: + +```text +ca_tti_score < 0.58 +or hag >= 0.42 +or trend_score <= 0.42 +``` + +In the implementation, `hag` was operationalized narrowly as a hallucination-amplification gap derived from unsupported claims, confidence-artifact mismatch, and missing evidence. In the broader paper model, this should be treated as one subtype of Human-Agent Alignment Gap rather than the full construct. + +## 6. Synthetic Results + +The first synthetic run used 8 trials and 10 steps per scenario, producing 400 trajectory-level observations. The results are summarized below. + +| Scenario | Failure Trials | Raw Warn Trials | CA-TTI Warn Trials | Mean CA-TTI Lead | +| --- | ---: | ---: | ---: | ---: | +| Artifact drift | 8/8 | 0/8 | 8/8 | 3.0 steps | +| Fluent hallucination | 8/8 | 0/8 | 8/8 | 1.5 steps | +| Clean baseline | 0/8 | 0/8 | 0/8 | n/a | +| Low-confidence good artifacts | 0/8 | 1/8 | 0/8 | n/a | +| Noisy interaction with stable artifacts | 0/8 | 7/8 | 0/8 | n/a | + +These results support the behavioral premise of CA-TTI. In the two failure scenarios, raw TTI did not warn at all, while CA-TTI warned in every trial. In the noisy but stable scenario, raw TTI produced warnings in most trials, while CA-TTI produced none. In the low-confidence good-artifact scenario, CA-TTI did not treat low confidence alone as failure when artifacts and evidence remained strong. + +The results should be interpreted cautiously. The generator was intentionally constructed to test known failure modes, and the scoring thresholds were not validated against real teams. The correct interpretation is not that CA-TTI is proven. The defensible claim is narrower: separating artifact score, confidence calibration, trend, and alignment-gap signals can make failure modes visible that a single raw score misses. + +## 7. Governance Interpretation + +CA-TTI should be used as a team-level diagnostic and early-warning framework. It should not be used to score individual developers, rank teams, or evaluate personal performance. Such use would distort behavior and create surveillance risk. + +The governance rule is: + +```text +A team should not be considered more transparent if artifact transparency improves while psychological safety, trust, or human-agent alignment declines. +``` + +This rule matters because AI agents can increase the quantity and polish of project artifacts. More complete artifacts are not inherently harmful. The risk is that the organization may mistake artifact fluency for shared understanding. CA-TTI is designed to prevent that collapse of interpretation. + +Practical deployment would require safeguards: + +1. Report CA-TTI at the team level, not the individual level. +2. Separate artifact score from confidence and HAG. +3. Preserve provenance for AI-created or AI-modified records. +4. Require explicit human confirmation for high-impact decisions. +5. Treat psychological safety and workload as hard governance constraints. +6. Allow participants to challenge, correct, or reverse AI-mediated records. +7. Avoid manager-facing dashboards that expose individual prompt behavior. + +## 8. Discussion + +The shift from the original artifact-oriented TTI to CA-TTI changes the paper's contribution. The original contribution was a protocol for evaluating a conversational AI mediator. The revised contribution is a measurement model for human-AI software teams. + +This change makes the paper more durable. Tool designs will change quickly as AI agents evolve. A measurement problem will remain: how can teams know whether AI participation is improving shared transparency or merely improving the appearance of traceability? + +CA-TTI addresses that problem by refusing to compress transparency into one score. A high artifact score with low confidence should not be interpreted like a high artifact score with rich evidence. A high artifact score with rising HAG should not be interpreted as success. A temporary noisy transition should not be treated like confirmed transparency decline. These distinctions are the main value of the framework. + +The synthetic stress test also clarifies what future validation should test. A real-world study should not only ask whether CA-TTI increases during an intervention. It should ask whether CA-TTI predicts cases where teams report lower alignment, lower trust, or lower psychological safety despite improved artifact completeness. + +## 9. Limitations + +This draft has four major limitations. + +First, the synthetic data are not real Agile data. They test measurement behavior under designed scenarios but cannot establish construct validity. + +Second, the current prototype is trajectory-level rather than event-level. A fuller version should generate or collect event-level data with fields such as event type, source, actor type, linked artifact, confirmation status, delay, metadata completeness, review depth, autonomy level, correction events, trust signal, psychological safety signal, and missingness flag. + +Third, HAG is only partially operationalized in the current prototype. The implemented version focuses on unsupported claims and confidence-evidence mismatch. The broader construct should include review depth, confirmation debt, attribution ambiguity, correction/revert rate, trust decline, and psychological safety decline. + +Fourth, the paper still requires citation and claim verification before submission. Several references are inherited from the previous protocol draft and must be audited for claim-reference alignment. + +## 10. Future Work + +Future work should proceed in three stages. + +First, improve the synthetic generator so it produces event-level software-team records rather than only trajectories. This would make the metric easier to explain and closer to the eventual field setting. + +Second, conduct coder-based plausibility review. Human reviewers should inspect generated scenarios and judge whether the synthetic events plausibly represent healthy teams, documentation-only improvement, human-agent drift, noisy transition, and true transparency decline. + +Third, run a small field feasibility study. The goal should not be causal proof. The goal should be to test whether CA-TTI can be computed reliably, whether HAG can be coded consistently, and whether team members find the outputs meaningful rather than intrusive. + +## 11. Conclusion + +Human-AI software teams need a way to detect transparency drift before it becomes visible as project failure. Raw artifact scores are not enough because AI agents can improve the visible record while weakening shared human understanding. CA-TTI responds by separating artifact transparency, confidence, trend, and Human-Agent Alignment Gap. + +The first synthetic stress test supports the plausibility of this framing. CA-TTI detected artifact drift and fluent hallucination scenarios that raw TTI missed, while avoiding false warnings in noisy but stable conditions. These findings are preliminary, but they show why a single transparency score is not sufficient for human-AI software teams. + +CA-TTI should therefore be understood as an early-warning measurement framework. Its purpose is not to declare that a team is productive or well-managed. Its purpose is to ask whether the team still understands, reviews, confirms, and controls the work that humans and agents are producing together. + +## References + +Ball, K. (2021). *Electronic monitoring and surveillance in the workplace: Literature review and policy recommendations*. Publications Office of the European Union. https://doi.org/10.2760/5137 + +Cinkusz, K., Chudziak, J. A., & Niewiadomska-Szynkiewicz, E. (2025). Cognitive agents powered by large language models for agile software project management. *Electronics, 14*(1), Article 87. https://doi.org/10.3390/electronics14010087 + +Cleland-Huang, J., Gotel, O. C. Z., Huffman Hayes, J., Mäder, P., & Zisman, A. (2014). Software traceability: Trends and future directions. In *Future of Software Engineering Proceedings* (pp. 55-69). ACM. https://doi.org/10.1145/2593882.2593891 + +Shneiderman, B. (2020). Human-centered artificial intelligence: Reliable, safe & trustworthy. *International Journal of Human-Computer Interaction, 36*(6), 495-504. https://doi.org/10.1080/10447318.2020.1741118 + +Stray, V., Moe, N. B., & Bergersen, G. R. (2017). Are daily stand-up meetings valuable? A survey of developers in software teams. In H. Baumeister, H. Lichter, & M. Riebisch (Eds.), *Agile Processes in Software Engineering and Extreme Programming* (pp. 274-281). Springer. https://doi.org/10.1007/978-3-319-57633-6_20 + +Stray, V., Moe, N. B., & Sjoberg, D. I. K. (2020). Daily stand-up meetings: Start breaking the rules. *IEEE Software, 37*(3), 70-77. https://doi.org/10.1109/MS.2018.2875988 + +Stray, V., Sjøberg, D. I. K., & Dybå, T. (2016). The daily stand-up meeting: A grounded theory study. *Journal of Systems and Software, 114*, 101-124. https://doi.org/10.1016/j.jss.2016.01.004 + +Umar, M. A., Lano, K., & Abubakar, A. K. (2025). Automated requirements engineering framework for agile model-driven development. *Frontiers in Computer Science, 7*, Article 1537100. https://doi.org/10.3389/fcomp.2025.1537100 diff --git a/Aidaily_ca_tti_manuscript_revised_stage4.md b/Aidaily_ca_tti_manuscript_revised_stage4.md new file mode 100644 index 0000000..2825c2a --- /dev/null +++ b/Aidaily_ca_tti_manuscript_revised_stage4.md @@ -0,0 +1,335 @@ +# Transparency Drift in Human-AI Software Teams: A Confidence-Aware Team Transparency Index + +## Material Passport + +- Origin Skill: academic-pipeline +- Pipeline Entry Point: Stage 4 REVISE +- Origin Date: 2026-06-26 +- Prior Draft: `Aidaily_ca_tti_manuscript_draft.md` +- Review Package: `Aidaily_ca_tti_stage3_review_package.md` +- Verification Status: STAGE 4.5 FINAL INTEGRITY PASSED / CORRECTIONS APPLIED +- Document Label: CA-TTI revised manuscript +- Contribution Type: Conceptual measurement framework with synthetic stress-test evidence +- Source Materials: `Aidaily_final_manuscript.md`, `.context/ca_tti_synthetic_validation_plan.md`, `.context/ca_tti_synthetic_output_review.md`, San Diego synthetic experiment output, and Stage 3 review roadmap. + +## Abstract + +**Background:** Software teams increasingly coordinate work through human communication, project-management artifacts, code repositories, code reviews, and AI-generated updates. Coding agents and AI assistants can generate pull requests, issue comments, documentation, summaries, and decision records. These artifacts can make a project appear more complete and traceable even when human review, shared understanding, and accountability are weakening. + +**Objective:** This paper introduces CA-TTI, a confidence-aware Team Transparency Index for detecting transparency drift in human-AI software teams. Transparency drift is defined as a growing divergence between visible project artifacts and the human team's shared understanding, review depth, and ability to explain or control the work being produced. + +**Methods:** We revise an earlier Team Transparency Index based on coverage, consistency, consensus, timeliness, and completeness. CA-TTI preserves these artifact-oriented components but changes the output from a single score into a multi-signal framework: artifact score, confidence, trend state, and Human-Agent Alignment Gap (HAG). HAG is treated as a family of alignment indicators, not as a single validated latent variable. We evaluate initial measurement behavior using a synthetic stress test across five controlled scenarios. + +**Results:** In the synthetic prototype, raw TTI missed all warning cases in artifact drift and fluent hallucination scenarios. CA-TTI warned in all such trials, with mean warning lead times of 3.0 steps for artifact drift and 1.5 steps for fluent hallucination. In a noisy-interaction scenario without true artifact failure, raw TTI warned in 7 of 8 trials, while CA-TTI produced no warnings. Ablation checks indicated that removing trend logic substantially reduced warning lead time in the two failure scenarios. + +**Conclusion:** CA-TTI is proposed as an early-warning measurement framework, not as a validated productivity metric. Its central claim is that transparency in human-AI software teams must be interpreted as a multi-signal condition. Artifact completeness can improve while shared human understanding declines. Future work should implement event-level datasets, test inter-rater reliability, and validate the framework in real teams. + +**Keywords:** human-AI software teams; team transparency; AI agents; software traceability; Agile software development; shared mental models; measurement framework; transparency drift + +## 1. Introduction + +Software teams do not coordinate only through formal records. They rely on stand-up meetings, chat threads, issue trackers, pull requests, code reviews, release notes, documentation, and informal memory. These sources create a distributed picture of work. A decision may begin in a meeting, be clarified in chat, appear indirectly in a pull request, and never be reflected in the issue tracker. This fragmentation creates a persistent gap between what the team informally knows and what the project system formally records. + +Earlier versions of the Team Transparency Index (TTI) were designed to measure whether communication, project artifacts, and confirmed team knowledge were aligned. The original use case was an AI mediator for Agile teams: a conversational system would ingest stand-ups, chat, issue data, and Git metadata; extract candidate decisions and action items; detect mismatches; and request role-aware confirmation before writing back to the project record. In that framing, TTI was mainly an outcome measure for evaluating whether the mediator improved transparency. + +The rise of coding agents changes the measurement problem. AI systems no longer only summarize or remind. They can generate code, open pull requests, update issues, draft documentation, propose decisions, and produce fluent explanations. Evidence from AI pair-programming and agentic code-review research suggests that AI systems are becoming part of everyday software workflows, although their contributions, adoption patterns, and review needs differ from human work (Peng et al., 2023; Zhong et al., 2026). This can improve visible traceability, but it can also create a new kind of transparency failure. The artifact layer may become more complete while the team loses human understanding, review depth, accountability, or trust. + +This paper calls that failure mode **transparency drift**: a gradual divergence between visible project artifacts and the shared understanding of the human team. Transparency drift matters because artifact fluency can look like coordination. A project may show more comments, cleaner issue records, faster summaries, and more complete decision logs while team members are less able to explain why decisions were made, who endorsed them, whether AI-generated changes were deeply reviewed, or how errors should be corrected. + +The central contribution of this paper is CA-TTI, a confidence-aware measurement framework for human-AI software teams. The framework is not a validated field index yet. It is a conceptual and methodological proposal supported by synthetic stress-test evidence. The key shift is: + +```text +CA-TTI = artifact score + confidence + trend state + Human-Agent Alignment Gap +``` + +CA-TTI is not intended to replace human judgment or rank individuals. It is an early-warning framework for team-level inquiry. Its purpose is to show when an apparently orderly system may be drifting away from shared human understanding. + +## 2. Related Work and Positioning + +CA-TTI sits between four research streams: Agile coordination, software traceability, human-AI teaming, and governance of workplace monitoring. + +Agile software development depends on frequent communication and shared context. Daily stand-up meetings can support awareness, coordination, and monitoring, but prior work also shows that their value depends on team context and meeting quality (Stray et al., 2016, 2017, 2020). Agile transparency therefore cannot be inferred from the presence of ceremonies alone. + +Traceability research raises a parallel issue. Software traceability is valuable, but it is often created ad hoc and after the fact, which limits its practical benefit (Cleland-Huang et al., 2014). Socio-technical congruence research similarly links coordination needs to actual coordination patterns in software teams (Cataldo et al., 2008). These literatures show that software transparency is not only a documentation problem. It is a relationship between work dependencies, communication, records, and shared understanding. + +Code review adds a further signal. Code review is not merely a gate for defect detection. It also supports knowledge transfer, maintainability, and shared standards; developers judge review quality through factors such as feedback usefulness, clarity, and reviewer expertise (Kononenko et al., 2016). When AI agents participate in code review, the question is not only whether feedback is syntactically correct, but whether humans still understand, contest, and integrate that feedback in meaningful ways. + +Human-AI teaming research provides the theoretical bridge. Human-AI teams require more than a tool-user relationship; they require coordination around shared goals, roles, communication, trust, and team cognition (Berretta et al., 2023). Shared mental models are especially relevant because a team can coordinate effectively only when members maintain compatible expectations about tasks, roles, and system behavior (Andrews et al., 2023). Empirical work on human-agent teams also suggests that communication, explicitly shared goals, trust, and perceived team cognition shape performance and collaboration (Schelble et al., 2022). + +AI-supported project work adds another layer. Recent work on requirements engineering, cognitive agents, and Agile project-management support suggests that AI and machine-learning systems can extract structured requirements information and simulate or support Agile project-management roles (Cinkusz et al., 2025; Umar et al., 2025). Human-centered AI research warns that useful automation must preserve human control, safety, and trust (Shneiderman, 2020). Workplace surveillance research adds a governance concern: measurement systems can become instruments of monitoring and pressure if they expose individual behavior without appropriate safeguards (Ball, 2021). + +These strands suggest a measurement problem rather than only a tool-building problem. The question is not simply whether AI can produce more complete project artifacts. The harder question is whether a team remains genuinely transparent when AI participates in producing those artifacts. + +## 3. Construct Boundary: What CA-TTI Measures + +This paper uses transparency in a team-level, socio-technical sense. It does not equate transparency with explainability of an AI model or with the number of project artifacts. Four forms of transparency are relevant: + +| Transparency form | Question | Example signal | +| --- | --- | --- | +| Artifact transparency | Are project records linked, current, and complete? | Issue, pull request, commit, and decision-record links | +| Process transparency | Is it clear how work moved from discussion to decision to implementation? | Trace from meeting or chat decision to implementation artifact | +| Epistemic transparency | Can affected humans explain and challenge the work? | Human review depth, confirmation, correction, and shared explanation | +| Governance transparency | Are provenance, access, consent, and accountability clear? | AI provenance labels, appeal paths, access rules | + +CA-TTI is designed for the intersection of these forms. The original artifact-oriented TTI mostly measured artifact transparency. CA-TTI keeps that core but adds confidence, trend, and HAG so that artifact improvement is not mistaken for full team transparency. + +The intended use is a team-level diagnostic and audit protocol. It can feed a dashboard, but the dashboard is not the primary contribution. The primary contribution is an interpretation rule: CA-TTI warnings should trigger team inquiry, not sanctions or individual performance assessment. A team uses the index to ask: "Do our artifacts still reflect what we jointly understand, review, and control?" + +## 4. From Artifact Transparency to Confidence-Aware Transparency + +### 4.1 Original Artifact-Oriented TTI: Artifact Transparency + +The original TTI used five components: + +```text +TTI = 0.25*COV + 0.25*CON + 0.20*CSN + 0.15*TML + 0.15*CMP +``` + +| Component | Meaning | +| --- | --- | +| COV | Coverage: eligible communication-mentioned tasks or decisions linked to an issue, pull request, Git artifact, or decision record. | +| CON | Consistency: status, ownership, blocker, and decision claims match structured records or are explicitly reconciled. | +| CSN | Consensus: high-impact decisions meet predefined role-confirmation thresholds before writeback. | +| TML | Timeliness: documented updates occur soon enough after the relevant event. | +| CMP | Completeness: action items and decisions include required metadata such as who, what, and when. | + +This score is useful because it makes artifact transparency operational. However, it remains vulnerable to two problems. First, the score can become overconfident under sparse or uneven data. If only one source is available, a clean-looking score may reflect missing evidence rather than real alignment. Second, the score can reward documentation hygiene even when human alignment is weakening. AI agents may improve links, summaries, and completeness while reducing review depth or obscuring responsibility. + +### 4.2 CA-TTI: Multi-Signal Transparency + +CA-TTI keeps the five artifact components but changes the output format. Instead of returning a single score, it returns four signals: + +```text +artifact_score +confidence +trend_state +human_agent_alignment_gap +``` + +The artifact score answers whether visible project records are linked, consistent, confirmed, timely, and complete. Confidence answers how much evidence supports that score. Trend state answers whether the system is improving, stable, deteriorating, noisy, or insufficiently observed. The Human-Agent Alignment Gap answers whether AI-generated artifacts remain aligned with human understanding, review, endorsement, and control. + +This separation matters because the same artifact score can have different meanings. A high artifact score with high confidence and low HAG is a strong signal. A high artifact score with low confidence is fragile. A high artifact score with rising HAG is the core risk case: artifacts look clean while shared understanding deteriorates. + +### 4.3 Signal Flow + +```text +team events + -> communication, issues, pull requests, reviews, decisions, AI actions + -> artifact score / confidence / trend / HAG + -> warning state + -> team inquiry and governance response +``` + +CA-TTI therefore should not be read as a scalar ranking. It is a bundle of diagnostic signals. The warning state is useful only if the team investigates the underlying components. + +## 5. Human-Agent Alignment Gap + +The Human-Agent Alignment Gap (HAG) is the distance between what agents produce or record and what the human team has actually reviewed, endorsed, understood, or accepted. + +HAG is not an anti-AI measure. It does not treat agent participation as harmful by default. It asks whether AI contribution is growing faster than the team's capacity for review, confirmation, and shared understanding. + +HAG should be treated as a family of subdimensions: + +| HAG subdimension | Definition | Observable indicator | Current prototype coverage | +| --- | --- | --- | --- | +| Confirmation debt | AI-created or AI-modified claims lack explicit human confirmation. | Unconfirmed decision records, auto-updated tickets, unendorsed summaries | Partial | +| Review-depth gap | AI-generated work receives shallower review than comparable human work. | Review rounds, comment depth, test discussion, reviewer expertise | Not yet implemented | +| Attribution ambiguity | The team cannot tell whether a claim came from a human, an agent, or a mixed process. | Missing provenance label, unclear author chain | Not yet implemented | +| Explanation gap | Affected humans cannot explain why a decision or change was made. | Post-hoc explanation checks, reviewer challenge outcomes | Not yet implemented | +| Correction gap | AI-generated artifacts require more correction, revert, or clarification. | Reverts, follow-up corrections, rejected suggestions | Partial | +| Trust and safety divergence | Artifacts improve while trust, psychological safety, or willingness to challenge declines. | Team survey or retrospective signal | Not yet implemented | + +The current synthetic prototype does not implement full HAG. It implements a narrower `hag_proxy`, a hallucination-alignment proxy derived from unsupported claims, confidence-artifact mismatch, and missing evidence. This is a useful first stress-test component, but it should not be interpreted as a complete operationalization of human-agent alignment. + +## 6. Synthetic Stress-Test Design + +The present draft uses a synthetic stress test to evaluate whether CA-TTI behaves sensibly under controlled failure modes. This is not a field validation study. It is a measurement-behavior test. + +The prototype compared a raw TTI-like score against a CA-TTI score. The generator produced 8 trials and 10 steps for each of five scenarios, yielding 400 trajectory-level observations. Each row included: + +```text +scenario +trial_id +step +raw_tti +artifact_score +confidence +unsupported_claim_rate +evidence_coverage +actual_quality +failure_label +``` + +The scenarios were: + +| Scenario | Purpose | +| --- | --- | +| Clean baseline | Raw TTI and artifact-centered signals should agree. | +| Artifact drift | Artifact quality deteriorates before interaction quality visibly fails. | +| Fluent hallucination | Output remains fluent and confident while artifact quality and evidence coverage collapse. | +| Low-confidence good artifacts | Useful artifacts exist but system confidence is muted. | +| Noisy interaction with stable artifacts | Interaction-level noise should not be treated as artifact failure. | + +The prototype scorer computed trend from a four-step artifact-history window. A negative artifact slope reduced the trend score; fewer than four observations defaulted to a neutral trend score. The prototype `hag_proxy` combined unsupported-claim rate, confidence-artifact mismatch, and evidence-coverage gap: + +```text +hag_proxy = + 0.52 * unsupported_claim_rate ++ 0.30 * max(0, confidence - artifact_score) ++ 0.18 * max(0, 1 - evidence_coverage) +``` + +Confidence was calibrated downward when `hag_proxy` rose: + +```text +calibrated_confidence = confidence * (1 - 0.65 * hag_proxy) +``` + +The prototype CA-TTI score was: + +```text +ca_tti_score = + 0.45 * artifact_score ++ 0.20 * calibrated_confidence ++ 0.20 * trend_score ++ 0.15 * (1 - hag_proxy) +``` + +Warnings were generated when: + +```text +ca_tti_score < 0.58 +or hag_proxy >= 0.42 +or trend_score <= 0.42 +``` + +The raw baseline used the same numeric warning threshold for the raw TTI-like score: + +```text +raw_tti < 0.58 +``` + +These thresholds were chosen as prototype stress-test settings, not as validated field cutoffs. A deployed system would require calibration by domain, team workflow, and evidence availability. + +## 7. Synthetic Results + +The results are summarized below. + +| Scenario | Failure Trials | Raw Warn Trials | CA-TTI Warn Trials | Mean CA-TTI Lead | +| --- | ---: | ---: | ---: | ---: | +| Artifact drift | 8/8 | 0/8 | 8/8 | 3.0 steps | +| Fluent hallucination | 8/8 | 0/8 | 8/8 | 1.5 steps | +| Clean baseline | 0/8 | 0/8 | 0/8 | n/a | +| Low-confidence good artifacts | 0/8 | 1/8 | 0/8 | n/a | +| Noisy interaction with stable artifacts | 0/8 | 7/8 | 0/8 | n/a | + +In the two failure scenarios, raw TTI did not warn, while CA-TTI warned in every trial. In the noisy but stable scenario, raw TTI produced warnings in most trials, while CA-TTI produced none. In the low-confidence good-artifact scenario, CA-TTI did not treat low confidence alone as failure when artifacts and evidence remained strong. + +An ablation check using the same scored observations suggests that trend logic contributed substantially to early warning. Removing HAG did not change warning counts in this small prototype, but removing trend reduced lead time below the failure step in both failure scenarios. The check is preliminary because it was not pre-registered and uses the same synthetic data. + +| Failure scenario | Full CA-TTI lead | No HAG lead | No trend lead | Artifact-only lead | +| --- | ---: | ---: | ---: | ---: | +| Artifact drift | 3.0 | 3.0 | -0.25 | 1.0 | +| Fluent hallucination | 1.5 | 1.5 | -0.88 | 0.25 | + +These results support only a bounded claim: separating artifact score, confidence calibration, trend, and alignment-gap signals can make synthetic failure modes visible that a single raw score can miss. They do not establish construct validity, external validity, or real-world predictive accuracy. + +## 8. Use, Misuse, and Deployment Vignette + +CA-TTI should be used as a team-level diagnostic and early-warning framework. It should not be used to score individual developers, rank teams, or evaluate personal performance. Such use would distort behavior and create surveillance risk. + +The governance rule is: + +```text +A team should not be considered more transparent if artifact transparency improves while psychological safety, trust, or human-agent alignment declines. +``` + +Practical deployment requires safeguards: + +1. Report CA-TTI at the team level, not the individual level. +2. Separate artifact score from confidence, trend, and HAG. +3. Preserve provenance for AI-created or AI-modified records. +4. Require explicit human confirmation for high-impact decisions. +5. Treat psychological safety and workload as hard governance constraints. +6. Allow participants to challenge, correct, or reverse AI-mediated records. +7. Avoid manager-facing dashboards that expose individual prompt behavior. +8. Retain raw event data only as long as needed for team-level audit. +9. Treat warnings as inquiry triggers, not sanctions. + +A safe deployment vignette illustrates the intended use. During a sprint review, a team sees that artifact score has risen from 0.71 to 0.84 because more pull requests, issue links, and decision records are present. Confidence is moderate, but HAG is rising because several agent-generated summaries were accepted without review and developers cannot explain the rationale behind two decisions. The correct response is not to identify a low-performing developer. The correct response is a team inquiry: review which AI-generated records require confirmation, add provenance labels, ask affected developers to explain or revise the decision records, and adjust review policy for future agent-generated changes. + +This use case also shows how gaming should be handled. A team could try to increase confirmations without improving understanding. CA-TTI therefore should not reward confirmation count alone. Confirmation must be linked to review depth, provenance, correction rights, and the ability of affected humans to challenge the record. + +## 9. Discussion + +The shift from an artifact-oriented index to CA-TTI changes the paper's contribution. The original contribution was a protocol for evaluating a conversational AI mediator. The revised contribution is a measurement framework for human-AI software teams. + +This change makes the paper more durable. Tool designs will change quickly as AI agents evolve. A measurement problem will remain: how can teams know whether AI participation is improving shared transparency or merely improving the appearance of traceability? + +CA-TTI addresses that problem by refusing to compress transparency into one score. A high artifact score with low confidence should not be interpreted like a high artifact score with rich evidence. A high artifact score with rising HAG should not be interpreted as success. A temporary noisy transition should not be treated like confirmed transparency decline. These distinctions are the main value of the framework. + +The framework also clarifies how future validation should proceed. A field study should not only ask whether CA-TTI increases during an intervention. It should ask whether CA-TTI predicts cases where teams report lower alignment, lower trust, lower psychological safety, or weaker shared mental models despite improved artifact completeness. + +## 10. Limitations + +This paper has five major limitations. + +First, the synthetic data are not real Agile data. They test measurement behavior under designed scenarios but cannot establish construct validity. + +Second, the current prototype is trajectory-level rather than event-level. A fuller version should generate or collect event-level data with fields such as event type, source, actor type, linked artifact, confirmation status, delay, metadata completeness, review depth, autonomy level, correction events, trust signal, psychological safety signal, and missingness flag. + +Third, HAG is only partially operationalized in the current prototype. The implemented `hag_proxy` focuses on unsupported claims and confidence-evidence mismatch. The broader construct should include review depth, confirmation debt, attribution ambiguity, correction/revert rate, explanation gap, trust decline, and psychological safety decline. + +Fourth, the thresholds used in this prototype are not validated field cutoffs. They are stress-test settings. Real deployment would require calibration, sensitivity analysis, and stakeholder review. + +Fifth, the literature base is stronger than in the initial draft but still selective. A full journal submission should include a broader review of AI coding agents, team cognition measurement, software repository mining, and responsible workplace AI governance. + +## 11. Future Work + +Future work should proceed in three stages. + +First, improve the synthetic generator so it produces event-level software-team records rather than only trajectories. This would make the metric easier to explain and closer to the eventual field setting. + +Second, conduct coder-based plausibility review. Human reviewers should inspect generated scenarios and judge whether the synthetic events plausibly represent healthy teams, documentation-only improvement, human-agent drift, noisy transition, and true transparency decline. Inter-rater reliability should be reported. + +Third, run a small field feasibility study. The goal should not be causal proof. The goal should be to test whether CA-TTI can be computed reliably, whether HAG can be coded consistently, and whether team members find the outputs meaningful rather than intrusive. + +## 12. Conclusion + +Human-AI software teams need a way to detect transparency drift before it becomes visible as project failure. Raw artifact scores are not enough because AI agents can improve the visible record while weakening shared human understanding. CA-TTI responds by separating artifact transparency, confidence, trend, and Human-Agent Alignment Gap. + +The first synthetic stress test supports the plausibility of this framing. CA-TTI detected artifact drift and fluent hallucination scenarios that raw TTI missed, while avoiding false warnings in noisy but stable conditions. Ablation checks suggest that trend logic is important for early warning in the current prototype. These findings are preliminary, but they show why a single transparency score is not sufficient for human-AI software teams. + +CA-TTI should therefore be understood as an early-warning measurement framework. Its purpose is not to declare that a team is productive or well-managed. Its purpose is to ask whether the team still understands, reviews, confirms, and controls the work that humans and agents are producing together. + +## References + +Andrews, R. W., Lilly, J. M., Srivastava, D., & Feigh, K. M. (2023). The role of shared mental models in human-AI teams: A theoretical review. *Theoretical Issues in Ergonomics Science, 24*(2), 129-175. https://doi.org/10.1080/1463922X.2022.2061080 + +Ball, K. (2021). *Electronic monitoring and surveillance in the workplace: Literature review and policy recommendations*. Publications Office of the European Union. https://doi.org/10.2760/5137 + +Berretta, S., Tausch, A., Ontrup, G., Gilles, B., Peifer, C., & Kluge, A. (2023). Defining human-AI teaming the human-centered way: A scoping review and network analysis. *Frontiers in Artificial Intelligence, 6*, Article 1250725. https://doi.org/10.3389/frai.2023.1250725 + +Cataldo, M., Herbsleb, J. D., & Carley, K. M. (2008). Socio-technical congruence. In *Proceedings of the Second ACM-IEEE International Symposium on Empirical Software Engineering and Measurement* (pp. 2-11). ACM. https://doi.org/10.1145/1414004.1414008 + +Cinkusz, K., Chudziak, J. A., & Niewiadomska-Szynkiewicz, E. (2025). Cognitive agents powered by large language models for agile software project management. *Electronics, 14*(1), Article 87. https://doi.org/10.3390/electronics14010087 + +Cleland-Huang, J., Gotel, O. C. Z., Huffman Hayes, J., Mäder, P., & Zisman, A. (2014). Software traceability: Trends and future directions. In *Future of Software Engineering Proceedings* (pp. 55-69). ACM. https://doi.org/10.1145/2593882.2593891 + +Kononenko, O., Baysal, O., & Godfrey, M. W. (2016). Code review quality. In *Proceedings of the 38th International Conference on Software Engineering* (pp. 1028-1038). ACM. https://doi.org/10.1145/2884781.2884840 + +Peng, S., Kalliamvakou, E., Cihon, P., & Demirer, M. (2023). *The impact of AI on developer productivity: Evidence from GitHub Copilot*. arXiv. https://arxiv.org/abs/2302.06590 + +Schelble, B. G., Flathmann, C., McNeese, N. J., Freeman, G., & Mallick, R. (2022). Let's think together! Assessing shared mental models, performance, and trust in human-agent teams. *Proceedings of the ACM on Human-Computer Interaction, 6*(GROUP), Article 13, 1-29. https://doi.org/10.1145/3492832 + +Shneiderman, B. (2020). Human-centered artificial intelligence: Reliable, safe & trustworthy. *International Journal of Human-Computer Interaction, 36*(6), 495-504. https://doi.org/10.1080/10447318.2020.1741118 + +Stray, V., Moe, N. B., & Bergersen, G. R. (2017). Are daily stand-up meetings valuable? A survey of developers in software teams. In H. Baumeister, H. Lichter, & M. Riebisch (Eds.), *Agile Processes in Software Engineering and Extreme Programming* (pp. 274-281). Springer. https://doi.org/10.1007/978-3-319-57633-6_20 + +Stray, V., Moe, N. B., & Sjoberg, D. I. K. (2020). Daily stand-up meetings: Start breaking the rules. *IEEE Software, 37*(3), 70-77. https://doi.org/10.1109/MS.2018.2875988 + +Stray, V., Sjøberg, D. I. K., & Dybå, T. (2016). The daily stand-up meeting: A grounded theory study. *Journal of Systems and Software, 114*, 101-124. https://doi.org/10.1016/j.jss.2016.01.004 + +Umar, M. A., Lano, K., & Abubakar, A. K. (2025). Automated requirements engineering framework for agile model-driven development. *Frontiers in Computer Science, 7*, Article 1537100. https://doi.org/10.3389/fcomp.2025.1537100 + +Zhong, S., Noei, S., Zou, Y., & Adams, B. (2026). *Human-AI synergy in agentic code review*. arXiv. https://arxiv.org/abs/2603.15911 diff --git a/Aidaily_ca_tti_stage2_5_integrity_report.md b/Aidaily_ca_tti_stage2_5_integrity_report.md new file mode 100644 index 0000000..61717f5 --- /dev/null +++ b/Aidaily_ca_tti_stage2_5_integrity_report.md @@ -0,0 +1,83 @@ +# Stage 2.5 Integrity Verification Report + +Manuscript: `Aidaily_ca_tti_manuscript_draft.md` + +Date: 2026-06-26 + +Pipeline stage: Stage 2.5, pre-review integrity gate + +Verdict: PASS WITH CORRECTIONS APPLIED + +## Scope + +This audit checked reference existence, bibliographic metadata, citation presence, claim-reference fit, and synthetic-result consistency for the current CA-TTI draft. + +The audit does not establish real-world construct validity for CA-TTI. It only verifies that the current draft's cited sources, local experiment figures, and stated limitations are consistent enough to proceed to peer-review simulation. + +## Reference Audit + +| Reference | Status | Notes | +| --- | --- | --- | +| Ball (2021) | VERIFIED / CORRECTED | The report exists. The official JRC page and final PDF identify DOI `10.2760/5137`; the draft previously used another EU DOI variant, `10.2760/451453`. Updated to `10.2760/5137`. | +| Cinkusz et al. (2025) | VERIFIED | DOI, title, journal, volume, issue, and article number verified via MDPI/Crossref metadata. | +| Cleland-Huang et al. (2014) | VERIFIED / CORRECTED | DOI and proceedings metadata verified. Author name corrected from `Mader` to `Mäder`; proceedings venue clarified as ACM proceedings. | +| Shneiderman (2020) | VERIFIED | DOI, title, journal, volume, issue, and pages verified. | +| Stray et al. (2017) | VERIFIED | Springer page and Crossref metadata verify title, authors, pages, and DOI. | +| Stray et al. (2020) | VERIFIED | IEEE/Crossref metadata verify title, authors, journal, volume, issue, pages, and DOI. | +| Stray et al. (2016) | VERIFIED / CORRECTED | SINTEF and Crossref metadata verify title, authors, journal, volume, pages, and DOI. Author names corrected to `Sjøberg` and `Dybå`. | +| Umar et al. (2025) | VERIFIED / CORRECTED | Frontiers and Crossref metadata verify title, authors, volume, article id, and DOI. First author shortened to APA-style `Umar, M. A.`. | + +## Claim-Reference Fit + +| Draft claim area | Status | Notes | +| --- | --- | --- | +| Daily stand-up meetings support awareness/coordination but depend on context and quality | SUPPORTED | Stray et al. (2016, 2017, 2020) support the contextual and mixed-value framing. | +| Traceability is valuable but often ad hoc or after the fact | SUPPORTED | Cleland-Huang et al. (2014) directly supports this framing. | +| AI/ML can support Agile project-management roles and requirements extraction | SUPPORTED AFTER TIGHTENING | The draft sentence was narrowed so Cinkusz et al. supports agent simulation/project-management roles and Umar et al. supports automated requirements extraction. | +| Human-centered AI requires human control, safety, and trust | SUPPORTED | Shneiderman (2020) supports the human-control and reliable/safe/trustworthy framing. | +| Monitoring systems can create workplace surveillance risk | SUPPORTED | Ball (2021) supports psychosocial and policy risks of workplace surveillance/monitoring. | + +## Data and Result Audit + +Local source checked: `/Users/trovo/conductor/workspaces/1/san-diego/experiments/ca_tti/sample_output` + +Files checked: + +- `manifest.json` +- `summary.json` +- `result_summary.md` +- `observations.csv` +- `scored_observations.csv` + +Verified values: + +| Scenario | Failure trials | Raw warning trials | CA-TTI warning trials | Mean CA-TTI lead | Mean HAG | +| --- | ---: | ---: | ---: | ---: | ---: | +| artifact_drift | 8/8 | 0/8 | 8/8 | 3.0 | 0.21 | +| fluent_hallucination | 8/8 | 0/8 | 8/8 | 1.5 | 0.369 | +| clean_baseline | 0/8 | 0/8 | 0/8 | n/a | 0.065 | +| low_confidence_good_artifacts | 0/8 | 1/8 | 0/8 | n/a | 0.059 | +| noisy_interaction_stable_artifacts | 0/8 | 7/8 | 0/8 | n/a | 0.094 | + +The draft table matches `summary.json` and `result_summary.md`. The sample output contains 400 observations plus one header row in each CSV. + +## Corrections Applied + +- Updated Material Passport verification status and version label. +- Replaced Ball DOI with the official JRC final-PDF DOI `10.2760/5137`. +- Corrected `Mader` to `Mäder`. +- Corrected `Sjoberg` and `Dyba` to `Sjøberg` and `Dybå` in the 2016 reference. +- Normalized the Umar first-author initials. +- Narrowed the AI/ML support claim so it matches the cited papers more closely. +- Added the 400-observation count to the synthetic result description. + +## Residual Risks + +- The manuscript remains a framework paper with synthetic stress testing, not a validated empirical study. +- HAG is conceptually broader than the current prototype implementation, which only captures a hallucination/evidence mismatch subtype. +- The literature base is still lean. Before journal submission, add more direct human-AI teaming, AI code agent, and team cognition literature. +- No full APA formatting pass has been performed yet. + +## Gate Decision + +The Stage 2.5 integrity gate is clear enough to proceed to Stage 3 peer-review simulation. diff --git a/Aidaily_ca_tti_stage3_prime_rereview_package.md b/Aidaily_ca_tti_stage3_prime_rereview_package.md new file mode 100644 index 0000000..b444488 --- /dev/null +++ b/Aidaily_ca_tti_stage3_prime_rereview_package.md @@ -0,0 +1,84 @@ +# Stage 3' Verification Review Report + +Manuscript reviewed: `Aidaily_ca_tti_manuscript_revised_stage4.md` + +Revision package reviewed: `Aidaily_ca_tti_stage4_revision_package.md` + +Original review package: `Aidaily_ca_tti_stage3_review_package.md` + +Date: 2026-06-26 + +## Decision + +Minor Revision / proceed to Stage 4.5 final integrity. + +The Stage 4 revision substantively addresses the Stage 3 major-revision roadmap. No second major revision loop is required. The remaining issues are integrity and polish issues: new references added during revision require full verification, and the final manuscript should receive a formatting/citation pass. + +## Revision Response Checklist + +### Priority 1 — Required Revisions + +| # | Original Review Comment | Author's Claim | Response Status | Revision Location | Verified? | Quality Assessment | +| --- | --- | --- | --- | --- | --- | --- | +| SC-1 | Reframe the manuscript consistently as a conceptual measurement framework plus synthetic stress test. | The manuscript now states that CA-TTI is a conceptual measurement framework with synthetic stress-test evidence, not a validated field index. | FULLY_ADDRESSED | Material Passport; Abstract; Sections 1, 7, 12 | Yes | Verified. The revised draft explicitly names the contribution type and repeatedly limits claims to synthetic measurement behavior. | +| SC-2 | Define HAG as a construct with subdimensions; distinguish conceptual HAG from the current `hag_proxy`. | A HAG subdimension table was added and the implemented synthetic measure is renamed/explained as `hag_proxy`. | FULLY_ADDRESSED | Section 5; Sections 6 and 10 | Yes | Verified. HAG now has subdimensions, observable indicators, and prototype coverage status. The construct/proxy distinction is clear. | +| SC-3 | Add reproducible synthetic-method details, including generator assumptions, thresholds, baseline warning logic, and sensitivity/ablation discussion. | The methods now include row schema, scenario purposes, formulae, warning thresholds, baseline rule, and ablation leads. | FULLY_ADDRESSED | Sections 6 and 7 | Yes | Verified. The revised methods are reproducible enough for a reader to understand the stress-test logic without external code. The ablation is properly caveated as preliminary. | +| SC-4 | Expand related work enough to position CA-TTI against existing software engineering and human-AI teaming literature. | Related work now covers Agile coordination, traceability, socio-technical congruence, code review quality, human-AI teaming, shared mental models, AI pair programming, and agentic code review. | FULLY_ADDRESSED | Section 2; References | Yes | Verified. The added literature materially improves positioning. Final source verification is deferred to Stage 4.5. | + +### Priority 2 — Suggested Revisions + +| # | Original Review Comment | Response Status | Notes | +| --- | --- | --- | --- | +| P2-1 | Add an operational "Use and Misuse" subsection. | FULLY_ADDRESSED | Section 8 directly addresses use, misuse, safeguards, and warnings as inquiry triggers. | +| P2-2 | Add a deployment vignette showing how a team should respond to a CA-TTI warning. | FULLY_ADDRESSED | Section 8 includes a sprint-review vignette with rising HAG and safe team response. | +| P2-3 | Add a table mapping CA-TTI signals to observable data sources and missingness/confidence rules. | PARTIALLY_ADDRESSED | The transparency taxonomy and HAG table map constructs to indicators. Confidence/missingness rules are discussed but not yet formalized as a separate table. This is minor and can be polished later. | +| P2-4 | Add a concise statement of target use: dashboard, audit protocol, research instrument, or governance framework. | FULLY_ADDRESSED | Section 3 states that the intended use is a team-level diagnostic and audit protocol. | + +### Priority 3 — Nice to Fix + +| # | Original Review Comment | Response Status | +| --- | --- | --- | +| P3-1 | Add a figure showing CA-TTI signal flow. | FULLY_ADDRESSED | +| P3-2 | Tighten terminology around transparency, traceability, shared understanding, and alignment. | FULLY_ADDRESSED | +| P3-3 | Polish APA formatting and decide whether to preserve diacritics consistently. | PARTIALLY_ADDRESSED | + +## Commitment Ledger Verification + +| Concern ID | Commitment | Fulfillment Status | Verified? | Notes | +| --- | --- | --- | --- | --- | +| SC-1 | Reframe as conceptual framework plus synthetic stress test. | fulfilled | Yes | Present in Material Passport, Abstract, Introduction, Results interpretation, and Conclusion. | +| SC-2a | Define HAG with subdimensions. | fulfilled | Yes | Section 5 includes six subdimensions. | +| SC-2b | Distinguish HAG from `hag_proxy`. | fulfilled | Yes | Sections 5 and 6 explicitly distinguish full HAG from the implemented proxy. | +| SC-3 | Add generator assumptions, thresholds, baseline warning logic, and ablation discussion. | fulfilled | Yes | Sections 6 and 7 include formulae, warning rules, raw baseline, and ablation table. | +| SC-4 | Expand related work. | fulfilled | Yes | Section 2 and References add relevant software engineering and human-AI teaming sources. | +| SC-5 | Add Use/Misuse plus safe deployment vignette. | fulfilled | Yes | Section 8 added. | + +No `COMMITMENT_GAP` findings were identified. + +## New Issues Discovered During Revision + +| # | Type | Location | Description | Severity | +| --- | --- | --- | --- | --- | +| NEW-1 | Citation integrity | References | Seven new references were added during Stage 4. They have plausible metadata, but the final integrity gate must verify existence, metadata, and claim-reference fit. | Minor / Stage 4.5 required | +| NEW-2 | Methods polish | Section 3 / Section 6 | The revision maps constructs to indicators, but confidence and missingness could be expressed in a compact table before submission. | Minor | +| NEW-3 | Formatting | References | Diacritics and APA proceedings formatting should be normalized across the full reference list. | Minor | + +## Decision Rationale + +The major concerns from Stage 3 have been addressed. The revised manuscript now has a clearer contribution category, a stronger HAG construct definition, a more reproducible synthetic method section, broader literature positioning, and a practical governance section. These changes materially improve the manuscript and resolve the main reasons for the prior Major Revision decision. + +The remaining gaps do not require another substantive rewrite. They are appropriate for the next pipeline stage: final integrity verification and final formatting. + +## Residual Issues + +1. Complete Stage 4.5 final integrity verification for all references, especially new Stage 4 additions and arXiv preprints. +2. Optionally add a compact confidence/missingness table during final polish. +3. Normalize APA style and diacritics in the final formatted manuscript. + +## Stage 3' Checkpoint + +Stage 3' is complete. + +Next pipeline stage: Stage 4.5 FINAL INTEGRITY. + +Recommended mode: full reference, citation, claim-reference, and data verification on `Aidaily_ca_tti_manuscript_revised_stage4.md`. diff --git a/Aidaily_ca_tti_stage3_review_package.md b/Aidaily_ca_tti_stage3_review_package.md new file mode 100644 index 0000000..0b4faaf --- /dev/null +++ b/Aidaily_ca_tti_stage3_review_package.md @@ -0,0 +1,404 @@ +# Stage 3 Peer Review Package + +Manuscript: `Aidaily_ca_tti_manuscript_draft.md` + +Date: 2026-06-26 + +Pipeline stage: Stage 3, full peer-review simulation + +Editorial decision: MAJOR REVISION + +## Field Analysis Report + +### Paper Basic Information + +- Title: Transparency Drift in Human-AI Software Teams: A Confidence-Aware Team Transparency Index +- Full text length: approximately 3,045 words +- References: 8 +- Integrity status: Stage 2.5 pre-review audit passed with corrections applied + +### Field Analysis + +| Dimension | Analysis Result | +| --- | --- | +| Primary Discipline | Software engineering / human-AI software work measurement | +| Secondary Disciplines | Human-computer interaction, Agile project management, workplace governance | +| Research Paradigm | Conceptual framework with synthetic measurement stress test | +| Methodology Type | Measurement model proposal plus simulation-style synthetic validation | +| Target Journal Tier | Q2/Q3 specialized software engineering or HCI venue in current form; Q1 potential only after stronger validation and literature expansion | +| Paper Maturity | Revised early draft: coherent thesis and structure, but needs stronger construct definition and methods reporting | + +### Recommended Target Venues + +1. *IEEE Software* - strongest practical fit if framed as a human-AI software-team measurement framework. +2. *Information and Software Technology* - possible fit after a much stronger methods and validation section. +3. XP / Agile Processes in Software Engineering and Extreme Programming - good fit if positioned around Agile transparency and human-agent coordination. + +### Reviewer Configuration Cards + +#### Reviewer Configuration Card #1 + +Role: EIC + +Identity Description: Senior editor for an applied software engineering venue focused on socio-technical software practice, human-AI collaboration, and actionable engineering measurement. + +Review Focus: + +1. Whether the paper has a clear contribution beyond the original Aidaily mediator framing. +2. Whether the manuscript is positioned as a framework, method, or empirical study. +3. Whether the novelty is strong enough for an international software engineering audience. + +Will particularly care about: The paper must avoid overclaiming from synthetic data and must clearly state what readers can use after reading it. + +Possible blind spots: May underweight the need for deeper measurement-theory grounding. + +#### Reviewer Configuration Card #2 + +Role: Peer Reviewer 1, Methodology + +Identity Description: Research-methods reviewer specializing in measurement models, construct validity, simulation studies, and reproducibility in empirical software engineering. + +Review Focus: + +1. Construct validity of CA-TTI and HAG. +2. Adequacy and reproducibility of the synthetic stress-test design. +3. Whether the results support the stated claims. + +Will particularly care about: The distinction between a measurement-behavior demonstration and a validation study. + +Possible blind spots: May treat a conceptual paper as if it were a mature empirical study. + +#### Reviewer Configuration Card #3 + +Role: Peer Reviewer 2, Domain + +Identity Description: Software engineering researcher working on traceability, Agile coordination, and human-AI software development practices. + +Review Focus: + +1. Domain grounding in traceability, Agile transparency, human-AI software work, and coding-agent literature. +2. Whether CA-TTI fills a real gap in existing software engineering measurement. +3. Terminology precision around transparency, traceability, alignment, trust, and psychological safety. + +Will particularly care about: Whether the literature base is sufficient for a new index. + +Possible blind spots: May prefer established empirical software engineering conventions over an exploratory framework contribution. + +#### Reviewer Configuration Card #4 + +Role: Peer Reviewer 3, Cross-disciplinary / Practical + +Identity Description: HCI and responsible-AI governance scholar focused on human oversight, automation accountability, worker monitoring, and team-level AI adoption. + +Review Focus: + +1. Practical deployment risks of a transparency index in real organizations. +2. Governance and surveillance implications. +3. Whether the proposed signals would be meaningful and acceptable to teams. + +Will particularly care about: Whether the paper prevents misuse of CA-TTI as a productivity or surveillance score. + +Possible blind spots: May ask for governance depth beyond the manuscript's immediate scope. + +#### Reviewer Configuration Card #5 + +Role: Devil's Advocate + +Identity Description: Adversarial reviewer focused on core-argument fragility, overclaiming, hidden assumptions, and the strongest counter-narratives. + +Review Focus: + +1. Whether the paper proves anything beyond behavior built into the synthetic generator. +2. Whether HAG is a real construct or a renamed mixture of confidence, review debt, and trust. +3. Whether the paper's contribution survives if the synthetic results are removed. + +Will particularly care about: The risk that the paper is an attractive concept with insufficient independent evidence. + +Possible blind spots: May under-credit the value of a well-bounded conceptual framework. + +## EIC Review Report + +### Overall Recommendation + +Major Revision + +### Confidence Score + +4 + +### Summary Assessment + +The manuscript has a timely and potentially valuable premise: software teams using AI agents need a way to detect loss of real transparency before artifacts look visibly broken. The shift from a mediator-evaluation paper to an index/framework paper is the right strategic move. The paper's best contribution is the four-signal framing: artifact score, confidence, trend, and Human-Agent Alignment Gap. That framing is memorable and publishable if the authors keep the claim bounded. + +The current draft is not yet submission-ready. It reads as a strong position/framework note with a small synthetic demonstration, not as a fully validated measurement paper. The contribution is promising, but the target reader needs a sharper statement of what is new compared with traceability metrics, trust/oversight frameworks, and existing human-AI teaming work. The paper also needs a clearer use case: who computes CA-TTI, at what cadence, using what data, and what action follows a warning? + +### Strengths + +1. Clear reframing from tool evaluation to measurement framework. +2. Strong central problem: artifact fluency can mask declining shared human understanding. +3. Good restraint in the Results and Limitations sections: the paper explicitly avoids claiming real-world validity. +4. Governance section correctly blocks individual-level scoring and productivity ranking. + +### Weaknesses + +1. The manuscript's contribution category is underspecified. It alternates between conceptual framework, metric proposal, and synthetic validation. +2. The literature base is too thin for a new index. Eight references are not enough to position a construct across software traceability, team cognition, HCI, and responsible AI. +3. The paper does not yet define the intended operational workflow for CA-TTI deployment. +4. The title promises human-AI software teams, but the empirical material is synthetic and trajectory-level, not team-level. + +### Questions for Authors + +1. Is CA-TTI intended as a dashboard metric, an audit protocol, a research instrument, or a governance framework? +2. What is the smallest real dataset on which CA-TTI could be computed? +3. Which decision should a team make when artifact score is high but HAG is rising? + +## Methodology Review Report + +### Overall Recommendation + +Major Revision + +### Confidence Score + +4 + +### Summary Assessment + +The methodology is appropriate for an initial measurement-behavior stress test, but not yet adequate for a validation claim. The manuscript is careful about this boundary, which is a strength. However, the synthetic study still needs stronger reporting. The reader cannot fully evaluate whether CA-TTI outperforms raw TTI because the generator assumptions, raw-warning threshold, CA-TTI thresholds, scenario parameterization, and sensitivity to these choices are not described in enough detail. + +The core methodological issue is construct validity. CA-TTI is built from artifact score, confidence, trend, and HAG, but the relationship among those components is not fully justified. HAG is especially underdeveloped: the conceptual definition includes review depth, endorsement, correction, attribution, trust, and psychological safety, while the implemented variable captures a narrower hallucination/evidence mismatch. This is acceptable only if the paper explicitly labels the prototype HAG as a proxy or subtype and presents a roadmap for operationalizing the full construct. + +### Strengths + +1. The paper uses synthetic scenarios for the right purpose: controlled failure-mode testing. +2. The manuscript separates measurement behavior from field validity. +3. The Results section reports both true-warning and false-warning behavior. +4. The Limitations section accurately identifies trajectory-level data as a constraint. + +### Weaknesses + +1. The synthetic generator is underreported. The paper needs enough detail to reproduce the five scenarios without reading external code. +2. Threshold choices are not justified. The values `ca_tti_score < 0.58`, `hag >= 0.42`, and `trend_score <= 0.42` may be reasonable, but no sensitivity analysis is shown. +3. Raw TTI is treated as the baseline, but the baseline warning rule is not explained. +4. HAG is conceptually broader than the implemented proxy, creating a construct-measure mismatch. +5. No uncertainty reporting is included. With 8 trials per scenario, the paper should avoid any language that implies stable performance estimates. + +### Required Improvements + +1. Add a compact Methods subsection describing generator logic, scenario parameters, thresholds, and warning definitions. +2. Add a table mapping each CA-TTI construct to observable indicators and current prototype coverage. +3. Add a sensitivity or ablation paragraph: what happens if trend, confidence, or HAG is removed? +4. Rename implemented `hag` in the prototype description to `hag_proxy` or `hallucination_alignment_proxy`. + +## Domain Review Report + +### Overall Recommendation + +Major Revision + +### Confidence Score + +4 + +### Summary Assessment + +The paper identifies a real software engineering problem: traceability and artifact completeness can improve while actual team understanding declines. This is a strong domain contribution if developed properly. The paper's connection to software traceability and Agile stand-up literature is plausible, and the the original artifact-oriented TTI components are intuitive. + +The domain grounding is currently too shallow. A new "Team Transparency Index" must be positioned against several literatures: traceability information models, socio-technical congruence, coordination breakdowns, shared mental models, team cognition, code review quality, AI pair-programming/coding-agent studies, and human oversight of AI-generated work. The paper does not need to cite everything, but it needs enough of this map to show that CA-TTI is not simply rebranding traceability plus trust. + +The term "transparency" also needs tighter treatment. In software engineering, transparency may mean visible work state, trace links, accountability, explainability, or shared situational awareness. CA-TTI uses all of these senses. That breadth is productive, but only if the manuscript defines the construct boundaries. + +### Strengths + +1. The artifact components of the original artifact-oriented TTI are easy to understand and domain-relevant. +2. The paper correctly identifies AI-generated artifact fluency as a new failure mode. +3. The distinction between artifact score and HAG is important and worth preserving. +4. The paper's governance stance is aligned with responsible software engineering practice. + +### Weaknesses + +1. Literature coverage is insufficient for a framework paper. +2. The paper needs a stronger taxonomy of transparency types: artifact transparency, process transparency, epistemic/shared-understanding transparency, and governance transparency. +3. HAG needs clearer boundaries from trust, psychological safety, and human oversight. +4. The original Aidaily mediator context is mentioned, but not enough is said about how CA-TTI generalizes beyond that origin. + +### Required Improvements + +1. Add a related-work matrix with columns for traceability, Agile coordination, human-AI teaming, AI coding agents, and governance. +2. Define "transparency drift" more formally, including what counts as drift and what does not. +3. Add a paragraph explaining why CA-TTI is not merely a traceability completeness score. +4. Expand HAG into measurable subdimensions. + +## Perspective Review Report + +### Overall Recommendation + +Major Revision + +### Confidence Score + +3 + +### Summary Assessment + +From a responsible-AI and organizational-governance perspective, the manuscript is strongest when it treats CA-TTI as an early-warning conversation starter rather than a managerial score. The governance rule is one of the most important sentences in the draft: a team should not be considered more transparent if artifacts improve while psychological safety, trust, or alignment declines. + +The practical risk is that the index could be adopted exactly in the way the authors warn against: as a dashboard used by managers to monitor teams. The paper names that risk but should go further. It should specify guardrails: aggregation level, access control, consent, opt-out, dispute mechanisms, data retention, and non-use for performance management. It should also explain how teams act on warnings without blaming individuals or discouraging useful AI assistance. + +The paper would benefit from a short deployment scenario. For example: during a sprint review, the team sees high artifact score but rising HAG; the response is a review-depth retrospective, not a productivity intervention. + +### Strengths + +1. Strong awareness that metrics can become surveillance tools. +2. Good separation between diagnostic use and productivity scoring. +3. The framework is practically understandable for engineering teams. +4. The paper keeps human confirmation central. + +### Weaknesses + +1. Governance safeguards are listed but not operationalized. +2. Psychological safety appears as a governance constraint but not as a measurable or procedural component. +3. The paper does not address stakeholder differences: developers, team leads, product managers, security reviewers, and executives may interpret warnings differently. +4. The manuscript should explain how to prevent gaming: teams might optimize visible confirmations without improving actual understanding. + +### Required Improvements + +1. Add a "Use and Misuse" subsection. +2. Add a short deployment vignette showing a safe response to a CA-TTI warning. +3. Specify who should and should not see the raw component signals. +4. Add explicit language that CA-TTI warnings trigger inquiry, not sanction. + +## Devil's Advocate Stress-Test Report + +### Critical or Major Challenges + +#### DA-1: The strongest result may be built into the synthetic generator + +Severity: Major + +The manuscript reports that CA-TTI detects artifact drift and fluent hallucination while raw TTI misses them. That result is plausible, but a skeptical reviewer will ask whether the synthetic generator was designed so that CA-TTI's exact components move before failure. If so, the result demonstrates internal consistency, not comparative detection power. + +Required response: The authors should describe the generator neutrally, include an ablation or sensitivity check, and state that the result is a behavioral plausibility test rather than evidence of superiority. + +#### DA-2: HAG may collapse several constructs into one label + +Severity: Major + +HAG includes review depth, human endorsement, attribution, correction/revert behavior, trust, and psychological safety. These may not move together. A team can have high trust but shallow review; strong review but low psychological safety; clear attribution but weak shared understanding. Treating them as one gap risks conceptual overreach. + +Required response: Split HAG into subdimensions or clearly present it as a family of alignment-gap indicators. + +#### DA-3: The paper claims to address human-AI software teams but has no human-team data + +Severity: Major + +The title and framing invoke human-AI teams, but the current evidence is synthetic and does not include real human review behavior, trust reports, psychological safety measures, or field artifacts from actual teams. + +Required response: Retain the title only if the abstract and conclusion clearly signal "framework and synthetic stress test." Otherwise, retitle toward "A Framework for..." rather than implying empirical field evidence. + +#### DA-4: CA-TTI could worsen the problem it aims to solve + +Severity: Major + +A transparency index can become a target. If teams know that confirmation, links, and timeliness raise scores, they may create ritualized confirmations and cleaner artifacts without deeper understanding. This is the same artifact-fluency problem at the metric level. + +Required response: Add anti-gaming and qualitative follow-up procedures. The index must trigger reflective review, not become a standalone score. + +### Strongest Counter-Argument + +The paper's strongest counter-narrative is: "This is not yet an index; it is a promising checklist with a synthetic demonstration." To overcome that critique, the revision must show a clearer construct model, a reproducible stress-test method, and a bounded claim. + +## Editorial Synthesis + +### Reviewer Summary Matrix + +| Dimension | EIC | R1 Methodology | R2 Domain | R3 Perspective | Devil's Advocate | +| --- | --- | --- | --- | --- | --- | +| Recommendation | Major Revision | Major Revision | Major Revision | Major Revision | Major issues | +| Confidence | 4 | 4 | 4 | 3 | n/a | +| Main strength | Timely reframing | Appropriate synthetic stress-test boundary | Real software engineering problem | Strong governance instinct | Useful concept survives if bounded | +| Main weakness | Contribution category unclear | Construct validity and underreported simulation | Thin literature grounding | Misuse/governance not operationalized | Synthetic result may be self-confirming | + +### Consensus Findings + +#### SC-1: Contribution type must be clarified + +Disposition: CONSENSUS-4 + +All reviewers converge on the need to identify the paper as a conceptual measurement framework with synthetic stress testing, not an empirical validation study. + +Required action: Revise the Abstract, Introduction, Methods, and Conclusion so the claim is consistently bounded. + +#### SC-2: HAG requires sharper construct definition + +Disposition: CONSENSUS-4 + +All reviewers identify HAG as valuable but underdefined. The current implemented proxy is narrower than the conceptual definition. + +Required action: Add a HAG construct table with subdimensions, indicators, data sources, and current prototype coverage. + +#### SC-3: Synthetic methods need reproducible reporting + +Disposition: CONSENSUS-3 + +EIC, R1, and DA raise this directly; R2 and R3 do not dispute it. + +Required action: Add generator logic, warning thresholds, baseline rule, and sensitivity/ablation discussion. + +#### SC-4: Literature base must expand + +Disposition: CONSENSUS-3 + +EIC, R2, and R1 raise this directly; R3 partly corroborates through governance literature needs. + +Required action: Add literature in human-AI teaming, AI coding agents, team cognition/shared mental models, socio-technical congruence, code review quality, and responsible workplace AI. + +#### SC-5: Governance safeguards must become operational + +Disposition: CORROBORATED FINDING + +R3 and EIC raise this strongly, and DA adds the gaming/misuse concern. + +Required action: Add "Use and Misuse" plus a safe deployment vignette. + +### Editorial Decision Letter + +Dear Authors, + +Thank you for submitting the manuscript "Transparency Drift in Human-AI Software Teams: A Confidence-Aware Team Transparency Index." The reviewer panel found the paper timely, clear, and potentially valuable. The central idea that AI-generated artifact fluency can mask declining shared human understanding is compelling. The move from a single TTI score to a multi-signal framework is also promising. + +However, the paper requires major revision before it can be considered submission-ready. The current version is best understood as a conceptual framework with a synthetic measurement-behavior demonstration. It does not yet provide empirical validation of CA-TTI in real human-AI software teams. The reviewers agree that the manuscript should lean into that bounded contribution rather than imply a validated index. + +The most important revision needs are: clarify the contribution type, define HAG more rigorously, report the synthetic stress-test method reproducibly, expand the related work, and operationalize governance safeguards. Addressing these issues would substantially improve the paper's credibility while preserving its core insight. + +Decision: Major Revision. + +## Revision Roadmap + +### Priority 1: Must Fix + +1. SC-1: Reframe the manuscript consistently as a conceptual measurement framework plus synthetic stress test. +2. SC-2: Define HAG as a construct with subdimensions; distinguish conceptual HAG from the current `hag_proxy`. +3. SC-3: Add reproducible synthetic-method details, including generator assumptions, thresholds, baseline warning logic, and sensitivity/ablation discussion. +4. SC-4: Expand related work enough to position CA-TTI against existing software engineering and human-AI teaming literature. + +### Priority 2: Should Fix + +1. SC-5: Add an operational "Use and Misuse" subsection. +2. Add a deployment vignette showing how a team should respond to a CA-TTI warning. +3. Add a table mapping CA-TTI signals to observable data sources and missingness/confidence rules. +4. Add a concise statement of target use: dashboard, audit protocol, research instrument, or governance framework. + +### Priority 3: Nice to Fix + +1. Add a figure showing CA-TTI signal flow: events -> artifact score/confidence/trend/HAG -> warning state -> team inquiry. +2. Tighten terminology around transparency, traceability, shared understanding, and alignment. +3. Polish APA formatting and decide whether to preserve diacritics consistently across all references. + +## Stage 3 Checkpoint + +Stage 3 is complete. + +Next pipeline stage: Stage 4 REVISE. + +Recommended mode: targeted major revision using the roadmap above. diff --git a/Aidaily_ca_tti_stage4_5_final_integrity_report.md b/Aidaily_ca_tti_stage4_5_final_integrity_report.md new file mode 100644 index 0000000..d081521 --- /dev/null +++ b/Aidaily_ca_tti_stage4_5_final_integrity_report.md @@ -0,0 +1,131 @@ +# Stage 4.5 Final Integrity Verification Report + +Manuscript: `Aidaily_ca_tti_manuscript_revised_stage4.md` + +Date: 2026-06-26 + +Pipeline stage: Stage 4.5, final integrity gate + +Verdict: PASS WITH CORRECTIONS APPLIED + +## Scope + +This final integrity pass verified: + +- reference existence and bibliographic metadata for all 15 references; +- in-text citation coverage against the reference list; +- citation-context fit for the major claims in the manuscript; +- synthetic-result consistency against the local experiment outputs; +- internal consistency of the revised manuscript's contribution and limitations. + +The pass does not establish field validity for CA-TTI. It verifies that the manuscript's factual scaffolding is accurate enough to proceed to finalization. + +## Corrections Applied to Manuscript + +| Item | Issue | Correction | +| --- | --- | --- | +| Berretta et al. (2023) | Reference listed an extra author, `Wrede, B.`, not present in Frontiers/Crossref metadata. | Removed `Wrede, B.` from the reference. | +| Cataldo et al. (2008) | Venue was incorrectly listed as the 2008 ACM CSCW conference. | Corrected venue to *Proceedings of the Second ACM-IEEE International Symposium on Empirical Software Engineering and Measurement*. | +| Schelble et al. (2022) | Reference omitted page range. | Added `1-29`. | +| Zhong et al. (2026) | Reference listed incorrect authors. | Corrected to `Zhong, S., Noei, S., Zou, Y., & Adams, B.` based on arXiv metadata. | +| Material Passport | Status still marked final integrity pending. | Updated to `STAGE 4.5 FINAL INTEGRITY PASSED / CORRECTIONS APPLIED`. | + +## Reference Verification + +| Reference | Status | Verification source | Notes | +| --- | --- | --- | --- | +| Andrews et al. (2023) | VERIFIED | Crossref / DOI `10.1080/1463922X.2022.2061080` | Authors, title, journal, volume, issue, pages, and DOI match. | +| Ball (2021) | VERIFIED | DOI resolver / EU Publications Office DOI `10.2760/5137` | DOI resolves to EU Publications Office manifestation. | +| Berretta et al. (2023) | VERIFIED / CORRECTED | Frontiers / Crossref DOI `10.3389/frai.2023.1250725` | Correct author list has six authors; corrected manuscript. | +| Cataldo et al. (2008) | VERIFIED / CORRECTED | Crossref DOI `10.1145/1414004.1414008` | Venue corrected to ESEM 2008. | +| Cinkusz et al. (2025) | VERIFIED | MDPI / Crossref DOI `10.3390/electronics14010087` | Official MDPI page lists *Electronics* 2025, 14(1), 87, published 28 Dec 2024. Kept journal-year citation as 2025. | +| Cleland-Huang et al. (2014) | VERIFIED | Crossref DOI `10.1145/2593882.2593891` | Authors, title, proceedings, pages, and DOI match. | +| Kononenko et al. (2016) | VERIFIED | Crossref DOI `10.1145/2884781.2884840` | Authors, title, ICSE proceedings, pages, and DOI match. | +| Peng et al. (2023) | VERIFIED | arXiv `2302.06590` | Preprint status retained; authors and title match. | +| Schelble et al. (2022) | VERIFIED / CORRECTED | Crossref DOI `10.1145/3492832` | Added pages `1-29`; article identity verified. | +| Shneiderman (2020) | VERIFIED | Crossref DOI `10.1080/10447318.2020.1741118` | Metadata matches. | +| Stray et al. (2017) | VERIFIED | Crossref DOI `10.1007/978-3-319-57633-6_20` | Book chapter existence, authors, title, and pages verified. | +| Stray et al. (2020) | VERIFIED | Crossref DOI `10.1109/MS.2018.2875988` | Metadata matches IEEE/Crossref record. | +| Stray et al. (2016) | VERIFIED | Crossref DOI `10.1016/j.jss.2016.01.004` | Metadata matches; diacritics retained for `Sjøberg` and `Dybå`. | +| Umar et al. (2025) | VERIFIED | Crossref DOI `10.3389/fcomp.2025.1537100` | Authors, title, journal, volume, article id, and DOI match. | +| Zhong et al. (2026) | VERIFIED / CORRECTED | arXiv `2603.15911` | Author list corrected. Preprint status retained. | + +## Ghost Citation Check + +Result: PASS. + +- No dangling in-text citations were found. +- No orphan references remain after manual review of multi-citation groups. +- Multi-citation groups verified: + - `Peng et al., 2023; Zhong et al., 2026` + - `Stray et al., 2016, 2017, 2020` + - `Cinkusz et al., 2025; Umar et al., 2025` + +## Claim-Reference Fit + +| Manuscript claim | Cited source(s) | Verdict | Notes | +| --- | --- | --- | --- | +| AI pair-programming and agentic code-review work show AI entering software workflows, with review/adoption differences from human work. | Peng et al. (2023); Zhong et al. (2026) | SUPPORTED WITH PREPRINT CAVEAT | Both are arXiv preprints; manuscript uses them as emerging evidence rather than settled consensus. | +| Daily stand-ups support awareness/coordination but depend on context and quality. | Stray et al. (2016, 2017, 2020) | SUPPORTED | Claim is appropriately bounded. | +| Traceability is valuable but can be ad hoc and incomplete. | Cleland-Huang et al. (2014) | SUPPORTED | Claim fits source. | +| Socio-technical congruence links coordination needs and coordination patterns. | Cataldo et al. (2008) | SUPPORTED | Claim fits source after venue correction. | +| Code review supports knowledge transfer, maintainability, and shared standards beyond defect detection. | Kononenko et al. (2016) | SUPPORTED | Claim is aligned with code-review quality framing. | +| Human-AI teaming involves shared goals, roles, communication, trust, and team cognition. | Berretta et al. (2023); Andrews et al. (2023); Schelble et al. (2022) | SUPPORTED | The manuscript does not overstate these sources. | +| AI/ML systems can extract structured requirements information and simulate/support Agile project-management roles. | Cinkusz et al. (2025); Umar et al. (2025) | SUPPORTED | Claim remains correctly split across the two sources. | +| Human-centered AI requires preservation of control, safety, and trust. | Shneiderman (2020) | SUPPORTED | Claim fits source. | +| Workplace monitoring can create governance and surveillance risks. | Ball (2021) | SUPPORTED | Claim fits source. | + +## Data Verification + +Local data source checked: + +`/Users/trovo/conductor/workspaces/1/san-diego/experiments/ca_tti/sample_output` + +Files checked: + +- `manifest.json` +- `summary.json` +- `result_summary.md` +- `observations.csv` +- `scored_observations.csv` + +### Synthetic Results + +Verified values match the manuscript: + +| Scenario | Failure trials | Raw warning trials | CA-TTI warning trials | Mean CA-TTI lead | +| --- | ---: | ---: | ---: | ---: | +| artifact_drift | 8/8 | 0/8 | 8/8 | 3.0 | +| fluent_hallucination | 8/8 | 0/8 | 8/8 | 1.5 | +| clean_baseline | 0/8 | 0/8 | 0/8 | n/a | +| low_confidence_good_artifacts | 0/8 | 1/8 | 0/8 | n/a | +| noisy_interaction_stable_artifacts | 0/8 | 7/8 | 0/8 | n/a | + +### Ablation Results + +Verified values match the manuscript: + +| Failure scenario | Full CA-TTI lead | No HAG lead | No trend lead | Artifact-only lead | +| --- | ---: | ---: | ---: | ---: | +| artifact_drift | 3.0 | 3.0 | -0.25 | 1.0 | +| fluent_hallucination | 1.5 | 1.5 | -0.88 | 0.25 | + +## Internal Consistency + +| Check | Verdict | Notes | +| --- | --- | --- | +| Contribution category consistency | PASS | Manuscript consistently frames CA-TTI as a conceptual framework with synthetic stress-test evidence. | +| Construct/proxy distinction | PASS | Full HAG and `hag_proxy` are separated. | +| Claims bounded to evidence | PASS | The manuscript repeatedly states that field validity is not established. | +| Governance interpretation | PASS | Warnings are framed as inquiry triggers, not sanctions. | +| Data/table consistency | PASS | Tables match local output files. | + +## Residual Non-Blocking Notes + +- Peng et al. (2023) and Zhong et al. (2026) are arXiv preprints. They are acceptable as emerging-context references, but a final journal submission should add peer-reviewed coding-agent literature when available. +- Cinkusz et al. is cited as 2025 because the official MDPI citation line is *Electronics* 2025, 14(1), 87, although the page publication date is 28 December 2024. +- Final formatting should normalize APA capitalization and proceedings style, but no remaining issue blocks finalization. + +## Gate Decision + +The Stage 4.5 final integrity gate passes after corrections. The manuscript can proceed to Stage 5 finalization. diff --git a/Aidaily_ca_tti_stage4_revision_package.md b/Aidaily_ca_tti_stage4_revision_package.md new file mode 100644 index 0000000..47b1f26 --- /dev/null +++ b/Aidaily_ca_tti_stage4_revision_package.md @@ -0,0 +1,127 @@ +# Stage 4 Revision Package + +Manuscript revised: `Aidaily_ca_tti_manuscript_revised_stage4.md` + +Prior draft: `Aidaily_ca_tti_manuscript_draft.md` + +Review package: `Aidaily_ca_tti_stage3_review_package.md` + +Date: 2026-06-26 + +Decision addressed: Major Revision + +## Summary of Changes + +- Reframed the paper consistently as a conceptual measurement framework with synthetic stress-test evidence. +- Expanded related work from 8 references to 15 references. +- Added a construct-boundary section distinguishing artifact, process, epistemic, and governance transparency. +- Expanded HAG into six subdimensions and separated conceptual HAG from implemented `hag_proxy`. +- Added reproducible synthetic-method details: generator fields, scenario purposes, formulae, thresholds, and baseline warning rule. +- Added an ablation table showing the effect of removing trend, HAG, and artifact-only warning logic. +- Added a "Use, Misuse, and Deployment Vignette" section. +- Updated limitations to cover construct validity, event-level data, threshold calibration, and selective literature coverage. + +## Revision Tracking Table + +| ID | Issue Description | Reviewer Source | Type | Section | Resolution Summary | Location of Change | Status | +| --- | --- | --- | --- | --- | --- | --- | --- | +| SC-1 | Clarify contribution type as framework plus synthetic stress test | Consensus-4 | Major | Abstract, Introduction, Conclusion | Added explicit contribution type in Material Passport, Abstract, Introduction, Results interpretation, and Conclusion. | Abstract; Sections 1, 7, 12 | RESOLVED | +| SC-2 | Define HAG more rigorously and distinguish proxy from construct | Consensus-4 | Major | HAG | Added HAG subdimension table and renamed implemented synthetic measure as `hag_proxy`. | Section 5 | RESOLVED | +| SC-3 | Report synthetic method reproducibly | Consensus-3 | Major | Methods, Results | Added row schema, scenario table, formulae, threshold rules, baseline warning rule, and ablation results. | Sections 6 and 7 | RESOLVED | +| SC-4 | Expand related work and positioning | Consensus-3 | Major | Related Work | Added human-AI teaming, shared mental models, socio-technical congruence, code review quality, Copilot, and agentic code-review references. | Section 2 and References | RESOLVED | +| SC-5 | Operationalize governance safeguards and misuse prevention | Corroborated | Major | Governance | Added Use/Misuse section, safeguards, safe deployment vignette, and anti-gaming paragraph. | Section 8 | RESOLVED | +| P2-1 | Add signal-to-data mapping | Review Roadmap | Minor | Construct model | Added transparency form table and HAG observable-indicator table. | Sections 3 and 5 | RESOLVED | +| P2-2 | Add deployment vignette | Review Roadmap | Minor | Governance | Added sprint-review scenario showing a safe response to rising HAG. | Section 8 | RESOLVED | +| P3-1 | Add signal-flow figure | Review Roadmap | Minor | Framework | Added text-based signal-flow diagram. | Section 4.3 | RESOLVED | +| P3-2 | Tighten terminology | Review Roadmap | Minor | Whole manuscript | Added explicit transparency taxonomy and bounded terms. | Sections 3 and 4 | RESOLVED | + +## Commitment Ledger + +```yaml +- concern_id: SC-1 + commitment_extracted: + - commitment_text: "Reframe the manuscript consistently as a conceptual measurement framework plus synthetic stress test." + commitment_type: restructure + required_evidence_type: prose_edit + fulfillment_status: fulfilled + +- concern_id: SC-2 + commitment_extracted: + - commitment_text: "Define HAG as a construct with subdimensions." + commitment_type: add_clarification + required_evidence_type: new_table + fulfillment_status: fulfilled + - commitment_text: "Distinguish conceptual HAG from the current hag_proxy." + commitment_type: add_clarification + required_evidence_type: methods_paragraph + fulfillment_status: fulfilled + +- concern_id: SC-3 + commitment_extracted: + - commitment_text: "Add generator assumptions, thresholds, baseline warning logic, and sensitivity or ablation discussion." + commitment_type: add_analysis + required_evidence_type: new_table + fulfillment_status: fulfilled + +- concern_id: SC-4 + commitment_extracted: + - commitment_text: "Expand related work enough to position CA-TTI against existing software engineering and human-AI teaming literature." + commitment_type: add_citation + required_evidence_type: new_citation + fulfillment_status: fulfilled + +- concern_id: SC-5 + commitment_extracted: + - commitment_text: "Add Use and Misuse plus a safe deployment vignette." + commitment_type: add_clarification + required_evidence_type: new_section + fulfillment_status: fulfilled +``` + +## Response to Reviewers + +Dear Editor and Reviewers, + +Thank you for the detailed and constructive feedback. We revised the manuscript to clarify its contribution, strengthen construct definition, report the synthetic stress test more reproducibly, expand the related work, and operationalize governance safeguards. + +### Response to SC-1: Contribution type must be clarified + +Response: We agree. The revised manuscript now states that CA-TTI is a conceptual measurement framework with synthetic stress-test evidence, not a validated field index. + +Changes made: The framing was revised in the Material Passport, Abstract, Introduction, Results interpretation, Limitations, and Conclusion. + +### Response to SC-2: HAG requires sharper construct definition + +Response: We agree. We added a HAG subdimension table and explicitly distinguish conceptual HAG from the implemented `hag_proxy`. + +Changes made: Section 5 now defines confirmation debt, review-depth gap, attribution ambiguity, explanation gap, correction gap, and trust/safety divergence. + +### Response to SC-3: Synthetic methods need reproducible reporting + +Response: We agree. We expanded the synthetic methods section and added formulae, thresholds, baseline rules, and an ablation table. + +Changes made: Sections 6 and 7 now describe the row schema, scenario purposes, scoring formula, warning rule, raw baseline, and ablation leads. + +### Response to SC-4: Literature base must expand + +Response: We agree. We expanded the paper's positioning across Agile coordination, software traceability, socio-technical congruence, code review quality, human-AI teaming, shared mental models, AI pair programming, and agentic code review. + +Changes made: Section 2 and the References list were expanded from 8 to 15 sources. + +### Response to SC-5: Governance safeguards must become operational + +Response: We agree. The revised draft now includes an explicit Use/Misuse section, deployment safeguards, and a vignette showing that warnings should trigger team inquiry rather than sanctions. + +Changes made: Section 8 was added. + +## Residual Limitations + +- The manuscript still does not claim field validation. +- The ablation is a small secondary check on the same synthetic output, not an independent experiment. +- New references added in Stage 4 still require final Stage 4.5 integrity verification. + +## Stage 4 Checkpoint + +Stage 4 is complete. + +Next pipeline stage: Stage 3' RE-REVIEW, focused on whether the revision actually addressed the Stage 3 review roadmap. diff --git a/Aidaily_ca_tti_stage5_finalization_package.md b/Aidaily_ca_tti_stage5_finalization_package.md new file mode 100644 index 0000000..d8ea5a5 --- /dev/null +++ b/Aidaily_ca_tti_stage5_finalization_package.md @@ -0,0 +1,47 @@ +# Stage 5 Finalization Package + +Manuscript: `Aidaily_ca_tti_final_manuscript.md` + +Date: 2026-06-26 + +Pipeline stage: Stage 5 FINALIZE + +Status: FINALIZED + +## Deliverables + +| Deliverable | Path | Status | +| --- | --- | --- | +| Final Markdown manuscript | `Aidaily_ca_tti_final_manuscript.md` | Created | +| Cover letter | `Aidaily_ca_tti_cover_letter.md` | Created | +| PDF manuscript | `Aidaily_ca_tti_final_manuscript.pdf` | Created | +| DOCX manuscript | `Aidaily_ca_tti_final_manuscript.docx` | Created | +| LaTeX source | `Aidaily_ca_tti_final_manuscript.tex` | Created as a PDF build artifact | + +## Final Formatting Actions + +- Created a final manuscript copy from the Stage 4.5 integrity-passed manuscript. +- Updated the Material Passport to Stage 5 finalization status. +- Added final declarations: Data Availability, Funding, Conflicts of Interest, and AI Disclosure. +- Prepared a generic cover letter suitable for a framework paper submission. +- Rendered DOCX using the local `make-pdf` renderer. +- Rendered PDF via a generated LaTeX source and local `tectonic`, after the browser-based PDF renderer could not start because its Playwright browser cache was missing. + +## Final Quality Checklist + +| Check | Result | Notes | +| --- | --- | --- | +| Final integrity status present | PASS | Material Passport says Stage 4.5 final integrity passed. | +| AI disclosure present | PASS | Added under Declarations and cover letter. | +| Data availability statement present | PASS | Synthetic output availability stated. | +| Funding statement present | PASS | No external funding declared. | +| Conflicts statement present | PASS | No conflicts declared. | +| References retained | PASS | 15 verified references retained. | +| DOCX generation | PASS | Created with the local `make-pdf` renderer. | +| PDF generation | PASS | Created with `tectonic`; minor overfull-box layout warnings remain in long lines/tables. | + +## Residual Notes + +- Before external submission, replace `[Author Name]` in the cover letter and add target-journal metadata if known. +- The PDF build produced minor overfull-box warnings around long lines and tables. These are layout polish issues, not content or integrity issues. +- The manuscript remains a conceptual framework with synthetic stress-test evidence, not a field-validated empirical study. diff --git a/Aidaily_final_manuscript.docx b/Aidaily_final_manuscript.docx new file mode 100644 index 0000000000000000000000000000000000000000..3aab776faa7720a3ce32b4c7ee20159c25e78468 GIT binary patch literal 20501 zcmY(qV{m8P6D=Iuwlx#mb~3S@iJggU+qV6SZQHhOJGpuOx8ADvemT49eA(4&uYJ0E z^*Zv>pkQb~KtNDHJkV>}sq=pQB)~vGh@e0~s6TJDgl%mcjcpwDl->RrJLu54T3Z#x zkIVKmAc;PE1|3=_aTxm{NzH-BkUJ{y_0&;@Exr)PXsvad#wXIrQ{~!_a5`M`?qjD^ z1YE8XkliJb2Wl81-$bNW)1%#c3ui7E~iwDZD z9F&^z85mCmhp_68V`LSSqO~#XUK7<-;2oWxv<427s}7$asf(AvG~^pED@vaD6anRK zPk9Sm9=jApc|_;@Hq~laY=Xmrz9|TuM9(12-_j2_j@KP+ee>sEw!BR)ZqDi0wBC7Q580@d^)YLpox~)@VK=YlgmW zJ3VTg4Qw4Qfjz3~lmP((g8fehmJTuk@Rhe3_(I!n-p_==0&x=KdH()>gzo1)SBNN zN{=|Z+|;}8Yiq;dEXlxz2ZSvJDfByh*R+AF1o{_oLM;6|0BOIncpqSvtooH8Evh0x zg+meN7@|HgkS&FZ8uG?Ga3Oid9FscIF*DL)zK%Q5jXgt~YGKQI(Af#GNQkP}7u~-D zw=I5u#yJv(F#@@(zY0(UJZfm;i#;ZaW5Bbe5982D{9;9icU?1cXr0}j&o?EO7BVw9 zy64NmJ`P;^Zp}9yikbA7z_?UoACnrfAI*IvfrNg!J+bpW_UlVi>M6e}bB15ohwS9pB;5!0fEe4fe;F{#M0jO##@M`Onx{_KY zQ__H9K}9EqiX-lUGt+40DkE81>>`mlwq$7LSG9C3b{&_1Lw3+$=HgXINMk~Gh7hHr z?iyn}%<1R7n79O84!08TH zAMTyigb>947X<`t(W@z8ARq-05Fq6L4~74K{Fh$q@wnpghwt8^Smw+!OCVIabTqnR zo2q-F2F+Bp8jDr6ZmZBlpaP{RX{jB`$SmGdU!z|uZ+0=Eks^M0t2}dV~bT~7FrnFys42xh>5>Uj7=g2T@b(QK22U8{K|24jJW7w?>7UY$_PIW zr@zm)ud+=KD|OQH6&yqIy}hON=rT+`KVDk0X(y$Y3+%07-t2W%sW)f%D+X_+)YvcG zIjGevH&@Cgt?N1Yuw!y~{x$#V_Lxx}ba08-9h)kzqdSJDbJLwVb~f4gex)wozSnL& zNb)fD-m^2dcMqv6O>x$FZ{PI3d~L4S_`b6F^dQJkd5c!HPZ~$P9{a7Ob~v%X>h&ho z1?Q3(=>vJ$BHFU|+#&Rx_9dgbhJpC?r=5eg9{$T`^4m@C^L0CV^Lv{&<-4VOJ5f}r zqW!onTjPB*fKbNU5{G;KjIv#w4e=!7&{*O6%S3hi`vuFTihs7bnd|$icQozWh@;}2 zkRT*QU)Mvo`Y|%qS|@Eyx%sF!hp_xvWi_#$Nlm2HU1qtdp?*-v{IQOvXR~9LOD)BA zF+-{vu4)Ya`^EmRLjfR4x9nUB)A2g|#it;2e{nnfj*fk`Z>J({wh!Q%Beo@lajvD; zbpYl5BCjKR9{f}(8QpXM#l`ieJ8grZvWl^KUqI?Ygv%A%NCMrdV4TNfeHx=7nLmNy znFu+y{($vaXZAZDLZ*!91m`WMJjJn&SyV2MP%v*+h4eAm#wJYV zqQUw^Ys08DO_eQy9jnyX&Ge)RG$}1@aPqX?TzAUmgzip$kIC)G!Wh>qMKf-K4ZM?= z0C1F~yJNUP0Me;SZAg@L^0unT4*IC3W!}1s*o1wFq(R-OAe~UX=vXizZek_Tj9oh7 zXsFyi^u72{X6JonWPeAX;-aa&y$8HC@$Xpr*`>iX(v*L-a5()BP%V^Lc&tBr@_H~v)U1k3yiLZ5WGts zti>#HXR|zI9%v6EPHP|8P z$a<+nOK?)769GhQ-e{BgR0w;Pp4^JTT7H8em=gT#zRO-6ooF(1Qj60ZrLjnV(mu4a zOlENcX+#`5P#4_QBS(F&N861NXA_*RwstDEI#Naq|Hr;Azg=cj0GkoB+ML~~@syD) zj9Jqd6ysJg>8rk<&`sKTT>SCm?d{SqRv@P5sdCiT$Kb}t_)((s>ek>JodZLpg!~__ z%{;1IY(OD=rYBk${f10m`@|nz;Y&=im=iH0?^sTL(t5uuAX8ups4QBPII?b4nn25B zCDBn3gTk+^_^8<%xy^v8R@5?i+1gzD^3&&`&ie%@4&j~8YD_)Z3hmE|HXE0Vs6?;Uwy)pn&hDP zegxYPTvU^n_SdKI&$8{2WlW-f^m~83?Qk;+*h|$e}L1=T}+afdNY}q1GQ( z?9tSuTkbA=X20k`R)I?O^;dSwQFhoV5B7cZ+^ZxLVWaCC3?Kh<(h5ll(5^4Tz|L%C z(;7VamVZyboUXn*=%sZX6lzaC#J#E-*ZsQJ)%~&u`?VL5WwQm_XchJxY@|5Hec-OC z`f+f;M#kw+z^9A~b|mp-e0fo0x(y>i&j6@A(f z)W2!(=ul7>Pb<2~W^8&~mBzW-XJRedm2w-?NH$Q1#7Rr-eqemV!WkYS4BNENp`HsQeJ@LDiIvM1FYU|qIEt95-53Ns15w?Kz7(dh@(F3@Oip(Rq5Usu-n^8zuqo?9-6GAI*5u{bf7qRB(oJzw4{aXI zDvhl!HnofgHvLQVH$4V|?ECnD_Y6V%zF3Z~X9UXqVO_2%TGL^H7m!7Cw1}KZZIm+cEBfZ*FPkjIZCsaTu<%uvEM5ZF;l5tG^O)?kFB|M|#espAJ zv}RxH%te3n>ti}gQ8umkSx#|VeRo*OM_$!wpZI{IX^f=GjE1B56fZSdoJBNXA-5U9 z51ng8#^%Fr?jPUrfEPqG#R9>B5dDkflF|L>)=0rPJpao>ZDw6+v!J3Cz|~=hQc99? zKdta!bkzx=n>!T>D(+ZsnMQdad^T$f5iDvU10}3}i*YSHhLp3J>7lb!-(9Bv3`e|5 zy4tPU)1)I3PAKNUE`?>@5ec*KWSn(;#2dgK*n=qUa1Gt&$?Yf}<$6J28(&?48Q(87 z5XwT6BA2Q)VBNMuOUPE{PTH^hM>FAdB~h6^)!?tBWQ``QiC`g?nqxL{$>eE9-y-qL z2h3@4cqUt+kEdBOE66XIx`g|`fC$lu3*qaqCMXk7SR_TcZ}rpPOR~o^vHYD53U%lj zHO>;x$o91${bt}KuL{i%cWtWlaa3J()D#2wZ_VM?^3QNcZpi9rjfiIVATFbfXbI}> zpt_t&8*Ak>fzjgh)^_uA&bFN|(g=3azG;ZIu>iWKUb#ITgIz9`K@>`EST^BH<4YkF zgJ_^P;!MQ0CK~gyU0t(ePsch3cTMskcO*EvDYkz{;ZCw_*9}#+D~ZBL?jz8^BN!(( z$FDOvF2`WZGJi?1{K~ge5K#|L z(@bU_U7W;Cp$eeE4aFZz6C*o>LyqYANDSl`c`Hb19sSGrxq|XMXtInwdqWx$WbpQjdc*cz zu6B-j2Fm+$_ebkZG3ghc2V6UF5{4ml8lS})Vk;q1K#{4ZE^y-UU9PTHAMT@Er7csFOx-eC(>Vw6B9!bNzb-?X zm;2jRLrbJxtUOInQH<29M*->cG3}er-kP+eKW#Nod-(A?9=x*Ok%`V{g_tYBb!88<$9M zM0n^zM#-wFYtmEOIUWQRX$q3opMmg++<=55%Q{u=B8od-L;qCASw!-ny3m9EPWDC= z;xocqEn)80W8QsRqw?LFK*~4FEo#Y9!AA)ZP?R*5B6jZ z{Yz58xnbac^a^1xN2)I*wj8S{&so42*IvWr4r<^qAd+DQ-R4m6Z_${77oKE$~?z8)uz2C%=bEhrngI3e>*^xM5*3V7}^LDC3~ER6Py(X-Nu&a)xIF6%6;P zWq9AAhD-NLLyJ2mFGVqt#5E360sEg?TgFx!r8v^+U5R!f?P_Ru!2)#y^=FGl3SJZU z%H#h^laOvXdAiSXmiUG5Yp?hB$LTJGp|Gf(qIa}iy?CwW|8+Tb`MprKMf&HumS+N) zylW?QJCSvkb943C{gf{}p8#6PI-~e=ipGx;&CQcXP2umasyfTW#>7R>FUkYsW$CIf zr`e7d*(h;L!zC1>)C^IL+Om7C3|g{+OZVE)LHkMTUrFMVtuCD`3CY&*<2ksEWI#!| z;d0SL3`Z;m6ZDN#hy#tF32(*3WX+=#(}j!A+2(Ye9`O|6-t&_5Utc{W9&$*DE0S_y zn@SownSm1uOn=nc(r6Vb&&S3ziXN%q_h4|Lp%x%{o1VeqTJ%+2pZR0%!CvEp7hGp& zQ1NRgZtnJosGhQkpu*?B#pEAAw7a;hAv>foXX0!g0<2a%{3X~jEmpti#TZ{N+O65z zr@nSmuXzPmOxduvLzvG3D(bpR*@Fn4|%IcZh#a{2%9v6+cdT*@l z!32FW{c62wu4g<~h4WD5$!s1&_ltQS@P{EFb!vZtf!odMif-l@Dzck8RC8dB54jhglcA?@LqM)>0EY(j zfu&!T3lls`cX*ej(HT(r6LCx7N4GWt*y4m4Bv`5VX}UI-ezB&%S{Gtr1h(loJxr4~_am&L)mo7AXS_p2yNU> zTpGCK2Gq@x@a*W<UKr#YP*mRkA!`I`0ip1St3r?XtN806_x|Y z%-9Fci>?8;ZVAOk@y%;93F}`6P>{oRb(qKAA#?dESE*}6ZVZaW))>!pVwHUu^Jl|sT z*x?Pb?P=(jxfsORB91~G!W?J?9|+D#H#z~y?Hp!5E=TgY;b2@PX&fIGt8Gjqjkm`; zNz!fLK6L{vgc^!l&X1NgEBxVNo*zqp^(Ft9QR|PzWowtT(}^dA(hYlFB8Ouk4~zc` zH1!UE>3y9Bn!iNi_6YiJ>_JOV`-;ebnSU$|PdTq~zPDEx?g;=Min@Z8XsJuiES9!I zNg>wy4X$#~)GlNPj7Sbh1jyK@c~x+La^i37{*yf&5 zrM?xJ90s{tuX(#nI)3Zbzyr<^KyC?Jnpt_mONP^s`VUp3K|}T^wOeTQsWD8Ru&6j^f>?g?%@s)ynZk#rhGOznRv+z> zzlHfHSys>y@U)l5&C94?5JHABJ}E;t{3yLrm^o1rNKc4u*Fl_- zB(LHp@q25!wUQN9p$rAaJae{KaHIKqYyXKpLfUmdgN2=NDA@K&W_G)s-m#CyrQ->r zhRoozmfiBOArt$5uT?*u3B1g6r&B(w1yCP;fN?P;dMLzf4e(=v>x;WIxBtmzvWZfO zvUfeD-9NP{zRR9CC|kpV|4gI|ZMJjHd0~Jw_1Ix*S40U*$`5Cn25%whbdroFQt&7_ zdmIi}dtZV|w=xwOWpmnB`J4rdl{hoK14GTjQm*(#jUbvHv?`wb)A78!;=7f0{B;G%C6E}^6W0B zgE~M2>Q@jH=bx zw5u;^hOX`4#$lmk(OhM49fp4F|7`h(skr~+SGMX=xP`mJ2ZW8R!lH3xlficJ&eC5x zO&iOF?m=i@SF0?`mEQB{h3W?Zu}xDHyj*%4Nu$&ax)z#Y=}LhAFh=v;xpyf=`6Qz( zHi48w7reg+RMvY7-8e}BAe1$Ml$Fz;A#WL77O_d}#vc8+Oq^$!OlpGtYb2}WRgCwr zdbk!7=P~rHoG`aSN=Y~J8)sUn0ysSgyT1nNGQWq0v!NpX6)(whzTVRukkP2huNL#L z-$G18067QTgs@O>YKWMi9flE?Ji0vJPpnqocL2pM{ZGRbR$xae+ zxz|=!69K8w_6+89eEYuT7WrGhx&G1lYD(26_9;i}MRgC{H8rLbN)E^DQ(ePHyGTEI z^EU*19^sDu=GmCkd5bOQH;T|*&+<*fiI4r}?<+kkLr-m7x~~ZJByRiLKVB8qQ~J}J zGDLnOBnpo(Pl2F1$%^j2xhs7#f`nkLJLYH{4f@-qd`BKI*@u-zj} z_TOYlvfpFgUdYVc1nz{34$Lco0i3;XsZEOaP5Ke%@+Y0&U@~H7CnvNFziJGD+B|;E z-f*Ek-|EL|7`IM31-l->$h*MqTYkC+RC?FEbj>n&&d)DXssomFD$uB$y}@HxVkQ6d z;89ZkuE&nSk1;8QZv(afWt~(|o?K)*2yAqe0bGt%krqh8r0zIJh}u`UhpTwG*@BU% zKF9ztAL)F*zcJSNJB_*enY^eRm>0Cq(Ak@rxa#f;fJGcCbUxLffKjMhVIgcJ($OwV zo5E7jc>`KUNAeLrjQ?d_N@axI=o0a}v}!Yi@rsdiH2og4LO(4^4!<(+J>_ zS8(~33gKbOmCBmc1)Ax?v|b}4Afe_P%nJcBmJM!%JqbLM;PX*19vB;E|6;9={STt( zUW{M4(`dRA6nhb3#+^nd+|ll49Zwf2=1uBfa{Ykzi`MaaqN#DLFy8Zz`ZtYgbs5rr zqd5^>kHdN?tq2!Z>SegEF>p{L>(oa^wfN5LfSDEIXX}4O`GCcbrZfVkhzcEq9@KfLGjZdC0L29)=EGPh+ zDdB;Ube*F(#>I~gB`y$nx2Hk^X=%dpb!obrV79L_fj_hkbMgGmB-#R!@r{EzDe~Yc za6``7mOxs1OzYxQitnxIcvaFc_$cuyam{eSN1)$y8fztRqB6x&v&=hqjm$nRmN{;% zHp_bDBKvw1iOXFtGwqFEmUGhe3=n)+4cd~&nfsd~Mun+V*Qebty8kg)75xiAwOKU! z4WX~^pZ(d_6_p_| z15Cmj|4#if$b-}Dh4z$uY5O<{Ol!t?nsH9r2qo68K3C9w)#jJ2_4THkbn+nsgU*B_ z`@wM309b0xMjhZ^y&vjptZtyR0o@ro3cqExAaHQ4=&d23S z+Ls4VD93w$bRgHSyx7O3IuR(TIQtSvL-S5pQpks8)X_%&ugwf(H?gU0CiYi8N$J4na_Dw{|6v0XrC-g+FJA#z6i@~A(k?!(WW zS%J!pXq0o=%{thO_`swG8%IcTE?+>l>-UIg>oa=6&Z;K;Z&KD|1(syAj56><`owJb zHjZ*?fpaN_H)CbUY13v+~GQAdp5P~?otWjh~+MYKJbFe6W6ra~%+WbmO zNWkEYg`?NLA_ea{EOyVk*s??(C3IiZUq1Xf3NnZqUg5j)WB8UCxy}EksjU*Y>|vRu z027zxAY#JaR_(~S`q*{2J2t4;9^gL+>h_>|{;T9vi1z>O-g+Vn?K znkIC>VWD!C{8muvOd^qh^VL~SGh12 zm&5-GW5vd3uW8QF{-?^8o9R3gh}HsE1+Ofgw~AAuuIFr`_$@iV$d*09 zz7UA#PfDl|AB-p1K2|uYSEbD;)lEhwC#qr!q=5Vr^KD0wO_>z#Lpw3;2GfY-+F2qa zNut&A4Bt9bAg&e-1U^l75A3JOSDN*JA`q+%L5mAsK{D3tps+N^0Gl5@ZUmQ&m{g~y zWn+q5-1J9(6Nj>(lZFht#(x#sajgdaBM9SX{n#vO!dEW7ylhJt0nK%Jr2V)am7rC1 z0QENg{V^Tv=w$B*JP+?=OL;6#@AGppCK+63%cmb%2PmU!8z>>a2GhzaNR3Brs~1(>iVCe3LQjQVi@zuv$EFx$|N4aib71w=~gMCQCJNqG@_K_Ru=>@OED43q+z$84Q|o z9}-VU!-<3}Cc9j3Gw8wklEyic>9}|)PF=;@l$QkDE`G9Zq31kNn?=T)hKMmwsXpB$ zPR3{Z{$lirB>-L7%vqe~>6r93_#x0m?AmB+pU}>6i@a(Qb0jEYR!ZC~u?jY54%eIv z&7WdpEu(|=;#U6yfTxE0uzsjN$-ibA$g}9@2wsek#|F1O;4SN5VJ6E=K z9|cpn5b2%;z<3*WTXl!UzZScA&O=I9r#=!?7*Sa^AT@1WV&#Gs;y6zQrKCsAn{kCD zUQ4&=K8emP^P#}D_c2F91FYGBMH}m0fK+rI_2pTTy%XU+&!)JYN~4BERuo(b3Biof zg;`TB$s%q*yx7>PPizpA8&{GrpZQmb9r5|h^3+JPHq4Qv6GS5SLxU3TS4(~F?|LCo z+7g9

X7=){s+-Zd6w9*Mne%wNO0KE4*d&^82`}uQ>}tFCgz73`p#Zcp5T49 zJ8DGly`L(r2s`VLKtvo(ZYMBjC86^1M}7a5!$$at7kgzEY;XdZ5`^Ja#O%1rVv@R! zUY19WFN7gE<$1Sgh<|VfZCO%WH;>_zc~I5!UF$&JQwis=%E=VO{UIL|t zUMD1Tei9*!?T=-8-T`7x4N;sFIiLoo29q^Zkgj>w4tvnBP*?O*SI01G%G;88$w!+? z)ZNXzVnPG6I6P#wHJ4H@95&yI>PUen`qu#W@k4_g2=75LTWhizW!48nN?mRcG-ONr zELS=-de1o#kZLg}USKDs1j-)e8O$DEgyLR^LGX|KQol_N-Vn+F#U!Z5_ISf~sY{H3 zP9g7*K{uzt5PudNfz4N8={ZRUtAXpJ1_~D!H3S}*A*}}=!KHkK_lPh~v&u1y*@JE@ zv101f7-%jb)sATz9xy_8c2Iq8l0??3e(eE@$zPG91sBi*qSN)sn{=uuw&}%Q`GM31 z`0`R+3ZtL}Q1@C67`(D^@6{%EZMiO$Tap&-HCQkk+G2;7h{J*VENOsP=hti4Hk!zuI=(z)A#`b={Qp`5Dl0`uOTqyOO7T{pmRq}>PUB{+^y%L#zq)Z za6kB%MQx%W0uhv8QWg-gWCh6+JqH^_aBV%42jomTn=X>VU_mFEnEVxtO~R~N1$-8# z1^X%`H~Wl&w5G1N!h`82VHtDV#T#K{eBLCB0!sfsA3(yb#3yP%UJ0btv8`|JGz{wd zNyi)SI>L4-wIb!{{u%|rTo|NoKZ@)6u9lgdEnmg8m++X21TWTNs@JX5z7^`+5YF^Aai~0)GYJ zcvJaj4Y?^eV++D}aU>lPbfDZfl;zXn%8oSQUBbvDW0fN4iXVsK3#4Yn<;>1I6^_J_ z{-BSM`#p&W-a*PpZ2b1V+L?MBXhP-6nr8hr>Xrf841a3nn-Ee2Dt@%b7t;7%jko6q z0OC#p&(Wnxz?RNojHE;>hFZm?dXaX*-2s(Uj)W^8yP_BG#L2P@VU^G&8{QU?O&^g`WuZkayFhI z6qbxwgYgJ#6-diFbz+5%UF#ZgaKkg0SM(F&Q?V7^f|(X@RFn!CWeFEjH)1r&BgaRg82Hm6n~9pN~RAEGlPgm7j5=dyYSQCXyhlBGs84x^I0S&FL(u z7W*a!>Fh~^4kdB{Rd!0r{XiN9D(qPH6=$AEVtqD$zqwB$1lfO=lMMUshxj!&Qa59@ z=*J!)4sH|D)zljK9cQdjq=$*_aQ)6HT{AZ}-I2Lzr|1?WS<+UQyZRc}`{EQ(ngZ&i z1a(}5({{PYDE|QCQNF@;qy;Z+jc3?Hqe!R(exj&$T+;-pSgs56Our0Fz^tfgG@KDKLA|2 z=Va_}k3@t6r!hxDo#siO_noV+|61O?AJgQ|jPL_eD}Brf%GFaE6Af3g2`t}@_(DN^ zhsCIZRvowoMT850o&T|PWI_xodyeA^L35?{pnrw)EfJhq^~s`azOg>DuAV;xkLtg# zhLym1y&j-61kVJLhNjcc+)|9(1k*oDBrjya!AT$yO+*N(3H&j6*Hf(z5u56vl&-$^+l znLqDseJ_aPg$l-Lj`kj}7c1dCS4k-p6P*OCPY>8R%>MnTFPDYdVPn(E*R zE&kaoteE++SftK<1&~0>$m!o@pviz_T1_~#iuEsjNqf(qN^?_|QKBaES%wmG=&-h@ zSDJbZ6N8M&&5u`0=YFuNy_>WeGTSmbFyfHd0(4#?Ygr*osAE`)#D^XQ>_`2%J zX3Ix@x%R>EbN+kU$wnyl8#Cga@inx+E9YP!%&4(LqrJVze-r-ICdrYKZsP7CN}^9) z3UMcUu-8r4f1xX!3o2&5?9&CjI_n))U|zcB0VNrH$psl;+=#O0Jw9FR3G z%{2)5ccIAfu)ra46NZTEC;6#PA#WWQ!6~HRUWdvHsER@N-a*hpsMPat`Kz5j?hhop zWsH&Q)a;_u=LV46CAqCr8%N7?8;V#uwk%fNxhHIcL{-A#hX0Pw&BI@r(zi>ZIROg$(BX(tR@uU{Bu|xD{_&*u=$x&W zQy?CRx?j>N)j;-vLcwh|YL@J*o?1vxWzO>LX>sYQPf|V4te4PdL9B`rHr7^?HUS>| z4IoR;`o2}?g`6DV?{f0XcV{4?kFu!7Q^a9!ZuSq9;_`5%=exfEvl_1&oy4;}aYGMd zwqIEAto{slsJ}>cE*0s_nj2p$KSK3U+sx~TPa|>Dz?Y$)C0r?6|5P_Ai27PvD4_d^ zVZ53BxO+~uZpWWwF$Zs!i>zVl6~6VU^%@AoDso+el_|sbxXDwG#=PLo4@&HNDS%Z& zB1!AS|J*FOC)6@=yU>adB)w#t#~z$*_7nUX9mmLF%>9<3ZCf%2<~i;RM<_O`oHMDV zM$F7}xMJ|2Kd!`PlUFQCu+&R2{RJ3bX4H#`^W;GYtwuO(oaQCY!=8MH3ib|{cJ5?e zX#jn2%H8^-ZEtvzfZLwSxmjBORhM4wzcw=vcxaKZBlJU2?&}0`I?EdioZPT5B~ZNk zx_GQCOU$eMKu43M!>lqqf(qCpjmXJxQT6I(T31=PoU1I4B{v&18v(xk)qhh?roxf! zaOnZ#9p7NBP`hRBl!A`G3<5ZItr)Byq@lYh#oc~5k3X@;G!Tn zNaj+yk%1A5|6hT`=Urf(4aLl`i46V87HK0BRmiadFcr$+O+S?nDffskSXa01QP3^| zU*Dl&=hirOJ_Jb7b0M$xv095jvdXqa26LRAjx62aF@8}TJz6BABmD`jcy=MN)NfN{kO9_`z$(;xnvvsB z#8=`-no-Y26>VQ9+m0E`v5G77zpXg5>=oh-a%Q{N!n^P8mDLNS6=pVtc8TbG0x4{wjb3T zP{~H;>W+dSSX%v_0!GlFIV(+9Zo-&tHdbkhD0-^`%lw6Z;RH}9UBgA9#kM(~14+ac z3S&OBE=U%U?slbOx_St=ZalG;nBZW!hS~15Cy66281cBlE5%cEk?@hGu?pwPJHL{- zyF*y~&cpoN5(e)G4`Re|J!;X97W%3JJeC1|OEcXPmHp=-DQ8~hUh*rT5I3<0drbcX zx~wxW`n3$T&*kMcUWcbm7>!$o_$E8K!w-X%#eD2sRIJ)7UM(ndthe$#q41to~;f zhsb9Ist4Xh_TJ)0Fd)^F9m*%kQ?my1DU6+`dxOOq2gS?Krcc{XlaS zhvhD-N0<;ib5^auP;hiB#?7ZC;4?UbbZdDA>m~Igh@SSPw@z}6*z<+M5oPkQWS;A% z6WM2Jf}Fme{Mcy5cszL*X3*EM#XTJsmrSEK?62OV`J;R>|5_Kc{&$IR21S?KN=dn_2vEe zpmo~*O+PicONOD#7TP@}8eeThh22^kf>#ROmaw<`NktV05ws&WW$;gA^MPIb|QCJiiGeMZWT*NQ_8`wHX~?jZ7Y{^zQ2K!O3j zMpHwl-qtGSX0?$Lx0RnMh9Oup(W#=}Z2hI{Df6~E0*M7d=xeG;js?k0DtB->LmCk1 zDdkd3GYrAg1R>Y@G-FNxV3T98_TUSbKLM33N38s)ja|;R+VAE#D{9CKbRBh2qWNi_ zSKIm2T8Yt4X(vSknP(Tv9~u;ox9sqT%W<`n>|rnFzk^MeaQt{fP*`$R%52(diuhd# z)S*0sH1+OwFktP2JlyoyNoko|Z&waVWoiGcE^)WY2V%ZX4CS0~y|EQ~HT-@Udzd4R zBud=U<%LbRmk9Pt^j;h{A01+}%}j8JTk9wxYFGApDBvsQsgQ4e8-&Qpfa91HB29dr ze;REYPSQKA`uWaY4iHif+i2QPstd3rZ9YHhTP2{Tr~_4ZO{p5k`9TxNWb=AFt%#nT zXOZL1eQMD=4d#7pq_f}hcQ6{GeFFwHB>d*+I>n5<keHcv@C804HNGm6cm+${ za6-mExCfh7Lq9R$wtGX^&S)4h-X&M+j%b9zXh*%px4(`3wfeV1T@KiGvK4POb@rBj zR=a_3s8gYGUbKAk?0THmG7}Ruqdbt$ZCmYmu)kSlCD@66Y5XT8R-^NI!|M6bHsb+cLzG=7L4T|M(r%>ot1T?6h;m& z*PBfgcwnWuocf|^jwjm|Tt|8)(OqSGa)0Nbo;LmC)@4xDW<{QqjVp3-BpxXgbwVU> zj)fC%ZZNN8TF|i>cURU*djcvhyKu7-Cma#SQrG|?rzl@U#Qhs!1DG3?sF`KwJ8Ao z-?6)HnA6eFs~Cf1Yc@_L-fICay1X717{w~%xtblWA&W1&W;Nky8IuOeAZxH+A3k&j%lLb;%DWcjuCahWLWotY9u?B z5cym{l9!Y%g`>z|lxGHox_RF@M%5C9{VUNwRwwrL%_WT$5yB5B2e75|gH2ylCW`5N z&0nuWPPSI8*R6XfZLp4Pe2j~+fAHgb@xY8FY_G=7q#A)He3#7Ch`K&mT#&nK=@0Mx zAzOvzP$-2AZW^Fuh3AGmVV8A5-kptcpRJAUPdVrfAD9^r3x!S#&L@w~!1#mNX6I^u z1{1n%==s3pw+B%+nReFpL(SqLdd9xg=NDCQmwKYhz`N2dX`G`q}ExjLs+=;L^TiGm{Sn4x2gtAQC@y}>@Y5T8IT&u^MKW_~;ndRG%7HXF)LkNf6p- zsQ7w@Zzc67%@s1U$7UkqlPjZP+86?gze5^@n>caMu0_Ma~8k!}0j9Ov%H#FxfH zP*$)kQKd>AX4shj&1QVa?B6Mg3^=6hndU8Vu>>NV+ZE(W?rmc&sUxw4-adYiyJp`WJiiqV#*B-n;s|B-_IXB z{rp2E|C>ubrC_(Z0QmInNC|L!@Aq~t(OG~XkS`Y{jX#!G0 z7ZHphq2?Fg@4ls%_1-&Y%{l+fcb_ved(U2L*35_G&_FR}u?1>YvNm+g?Jf@#5IdQp zMk69r8b>30z}6772%?ijoEanHDWt0<{w8ZPW9jXz6M(YcUWv{j3 z$0Aft88ny$(9@-syaM316*Guerpr5SCjG+vl}+a(%ILAqfr}w1Fug?MaW$>?J)H_u zW$&TJ`O|yP=_DlCyBLy-PNmUvFQtz+%&~I@U2QevE;V6W zu|A_Xn>(?T{PxbXxBl%A>U{adr$SM#(glXC-+gY@DSi2-m?h^K@5i>1dSQYisn&il zzZe1uAFnMYrI3&y2)chqx@Li~+*e>q>;t4acja#g{2>vSG(bt)<`6s^n9>vNfY6QT zApt$_mh#-Oe`Ujam2dGve8*km!IlkELKi^b(B`|V;*V3ev6%FT6k57fLuRjzl&jw& zso)|3-h@sgJRo#Hn#5m%!H+9N8I zd7i&CmZGn0TNB9=+*3*3dz_B~Cp9VCHO+5lSC5{BAtF}il?>eWzQ-}n#xL6~Ieo7h ze;Y)hot1TL2P~x)y!}Atd3OF#-(bd@kpQ2=ot5%vHqXhnOx+r;lBLCKZek6|TcdgB z#B56na%F0V>iR(t_z0-Dtqn;`sZOfkVj^rjml`>K> z59iUFRe5jHqh25|U36mRrerQ&U0Vwno`s^d2L&IVM!A0ePTQSbrLAeWI!Zmhu!g^- zauR|@%hW_NzKnJ*3{dE8TzhE!RbcVSEWMa^2G~~Ur|@nW;BRlC)f+-NcD~Epm%@}5 zwo8xdZWy?zx09}$?me6Jn0i)MYwBy3RUZg~=xoygf=J##AP_S4wSy)zJ6`C|ELgPB8F(H_hssOZ- z+ETOS+_%6o(;xVd9NMGpUY}2x^+mrP&zK1S^$qQMdRy*~&wIvd-z=(@Roqz$yNrM< zAi!5TD+8}H&t|MOs9JUPZFRCO!D*Cu>CI$)13Sd9XD9kv4i5w^iH}OuCbNp{D`{OM zfFIx1daf&scGK0;6V=}`0FmZuB^fKb-3v}l8TPZGebffgtsI@gZ=wz&h_~#+<^ih$b|{+$92jgRcY@@99lwd zb^}OIXH(iejPIKHI9AbC>0a?oR+%qB$_wukBX5*b<`jRs{0V8xQ5;F55D_U#E`Bw^nCBtsUSdFNa*-GaLivlc%h568G>FVC@&({ z*OjR1xyUXJrp9&TCGR(UTwX@g?+lEST$pw=TcR#!s3!{#`(rVp2%6rVb#K1h>$_-W z+Sucq>|*51R-fpr<-Lh}rZDRdemEFyObfP8|0n#zBW}KFib2=`F~}c)f5SgQZ+ZE9 zDgEnt_9p;k-q|vcbdI-?>cm_BtvgC9nAf`;dge#}36t4;@2mMT{--U(voyD;qP9xi z75v{iQcjc{J7*q4?oAPS6iuF&@m|W}%mPFweI=--o<<~VlJNZWw>$={g61X>R?aQU z{7rUAO%waU##l#-kmU0@AABp?N*S9az3}XrNxRojC|qTXL~y>lJio$;Ii@PD%A7I3 zfO%<>;GJ~lFQSu3P*H#NY+H@u&Vv@9Q^i&Ivj})qcO0Cp69J`2xy}E{@xyAA|9(-e zrr&Z(E1a)%x`m(cz!iwp2zwjiQKkSMD^O(|cqeRRLr0x1STp7~ zEc~8CWa2lOF4Jp`y5au%sx`C2vCfGvSq9_>5%;G=XMmWYZa41r00h;P=6;UasLfnI zLfR|e!@T}NiH@b--HlE-_WDFfAbcbb<+Xb!0sDQ-tLXkJ*X9@V2&VPCNhZB+m1-wN z!qD8!5BM6RgxLNu8F`XfLnafPskS#*en9L;>{?3Bx`jYpQQP^CmQQrARy=}}Q|Z!* zdgWeGWqWfQ#%|deB;k40wGVV%&oFT#z3Owld8BA^RI&h0;%kl(~SrR zQuq6ZTcb0?@j>(@ZCx8%&eAowm9N}hNdr%0U=!)CqX?MPoa7UV1_Zd%Q!YDqmcH20 zOZ(VnVc?GjK$4Rms_Ey6-sPvVtZuGM0=(@E;o}^kT8$%q^)vWD>z0jI<2grLyHRTD za>~JOI7)4$A%UCxl4Bg!51__lX>>QIvgejsVIGr)=fxO{oSRGhK%*C|4u7fi=G&ZK z%nOTi?U%!JOShFhg?C@xo)~%cX8}0aylkzKC5c=9xdrPBSXl=lyvwh%-pwVys4sM( zyJCNGz`MJ8X!*I*z6Yv$sE{L}Llo4j9rg}tP7_;;u21S){JNRvG#3L=niVpcs^vT( zraGhZ`s!j;=OfsrqYHmV{t6doxU8h<`+ca(gz={Bh^^!(JU4!U7M7y(MoT#dQ*tjr zc&J$XSr=6%8zMEjcHzy`Wh$v<6Vmnxrt@)_2EW>oor+G|w>k}-Iv-AslKL`I43z15 z7sJJJ3IlA|JMzyNVcfUw%16db;KzU>{loz=NI*8hfk7ck9)ZDLKW77f1_=pwp7$pS z=?&GBpG#80JP>_6;qv;so=3}rt4DM^i2HYI3=7;5xSbwc&$vG!K;9ZP$~J{5DswbE zb-(=H>d^RpX|w}t%BCr##r`$8%|~YAF1})y5CJGFR6=in(v*cS!KqoSIYC zy4Ef`{x;s@q=b*|%9H6vtXT7tTLmUKN9T^AlEw@+C$C*AqNFmr+7l@y?n^_>N(v~E zscX*Z28ETuNudIJ98TLkP}apr>_Tt+!j5nREb_DB>&#&{jFFx35jy9|#sgd^=k!~h z&lpW7GYaqTU2r2P=w}Z+@6+RIRX(0LGwH@xQB!(U^2Kf79cam#MBN@!NT3<^#U9m^ zWm5E*l5S3s%bxKWUZPdWeB;YPrmDG{B5&&{PiLh~{~ZZGg-`Dll?LN zg<1O1c;ukT0L2VZ-9KloOhj1zhtNJKKNtwV + + + + + + + + + + +

# Protocol for Evaluating a Conversational AI Framework for Agile Team Transparency and Knowledge Traceability

+


+

## Abstract

+


+

**Background:** Agile teams rely on daily coordination, issue tracking, and version control to maintain shared understanding. In distributed and hybrid teams, however, decisions, blockers, and action commitments often remain fragmented across meetings, chat, Jira, and Git. Conversational AI may help by extracting candidate updates from informal communication, linking them to project artifacts, and prompting team members to confirm or correct them.

+


+

**Objective:** This protocol describes a mixed-method field study to evaluate whether a conversational AI framework improves Agile team transparency, knowledge traceability, and perceived alignment without increasing cognitive burden or weakening psychological safety.

+


+

**Methods:** The study will use a two-stage mixed-method design. Stage A is a feasibility pilot that estimates measurement reliability, prompt burden, extraction/linking accuracy, and safety signals. Stage B is an optional powered field evaluation using a quasi-experimental, baseline-to-intervention design across Agile software teams. During baseline, teams continue normal workflows while communication, issue, and version-control metadata are measured. During intervention, teams use a conversational AI mediator that ingests stand-up transcripts, chat messages, Jira data, and Git metadata; extracts candidate decisions and action items; detects inconsistencies; and requests role-aware confirmation before writeback. The primary feasibility outcome is whether the Team Transparency Index (TTI) can be computed reliably. The primary effectiveness outcome, if Stage B proceeds, is change in TTI from baseline to intervention.

+


+

**Analysis:** Feasibility analysis will report reliability, prompt-burden rates, technical performance, missingness, and progression criteria. Effectiveness analysis will compare baseline and intervention periods using mixed-effects models or non-parametric alternatives when assumptions are not met. Qualitative interviews and observation notes will be analyzed thematically to explain adoption patterns, trust, interruption costs, and governance concerns.

+


+

**Ethics and Dissemination:** The study requires informed consent, role-based access controls, provenance-preserving audit trails, and safeguards against individual performance scoring. Results will be reported as a protocol-compliant field evaluation and will distinguish system performance from team-level organizational outcomes.

+


+

**Keywords:** Agile software development; protocol paper; conversational AI; mixed methods; field study; team transparency; knowledge traceability; psychological safety; human-AI collaboration

+


+

## 1. Introduction

+


+

Agile software development depends on shared context. Teams coordinate through daily stand-ups, chat threads, issue trackers, code reviews, and version-control activity. These artifacts are individually useful, but they do not automatically produce a durable and verified team memory. A decision may be made verbally, partly clarified in chat, reflected indirectly in a pull request, and never updated in the issue tracker. The result is a persistent gap between what the team knows informally and what the system of record says formally.

+


+

Prior research on daily stand-ups shows both the value and limits of recurring Agile communication. Stand-ups are widely used and can support team awareness, but their value varies by team size, role, and meeting quality (Stray et al., 2017). Grounded theory work also shows that stand-ups support coordination and monitoring while remaining sensitive to local practice (Stray et al., 2016). Later work argues that teams should adapt stand-up rules when the ritual no longer serves communication needs (Stray et al., 2020). These findings suggest that Agile transparency cannot be inferred from ceremony adoption alone.

+


+

Research on Agile scaling and methodology fit also supports a context-sensitive approach. Verwijs and Russo (2024) found that scaling frameworks themselves explain little practical difference in team effectiveness. Itzik and Roy (2023) argue that Agile fit depends on software project characteristics and should be assessed through a decision framework. For a transparency intervention, this means the target is not framework compliance but the quality of alignment among people, artifacts, and decisions.

+


+

AI-supported software project management provides relevant technical foundations. Automated requirements engineering research shows that machine learning can extract structured models from natural-language requirements in Agile contexts (Umar et al., 2025). LLM-based multi-agent project-management frameworks such as CogniSim show that AI agents can support Agile roles and project workflows in simulated environments (Cinkusz et al., 2025). Comparative work on Agile methods and technology-enhanced practices also suggests that AI-enabled tools may affect delivery and quality outcomes, while introducing risks of over-reliance on automation (Malla, 2025). The traceability literature is also directly relevant: software traceability is valuable but often performed ad hoc and after the fact, limiting its realized benefit (Cleland-Huang et al., 2014). Human-centered AI research further emphasizes that reliable and trustworthy systems should combine high automation with meaningful human control (Shneiderman, 2020). These strands motivate a field study that tests whether conversational AI can improve the human communication layer of real Agile teams without producing automation bias, intrusive monitoring, or reduced psychological safety (Parasuraman & Riley, 1997; Ball, 2021).

+


+

This protocol defines a study to evaluate a conversational AI mediator for Agile transparency. The protocol is informed by general protocol-reporting principles for transparent intervention studies, including the logic of SPIRIT-style completeness and AI-specific reporting attention to intervention behavior, human oversight, and error handling (Chan et al., 2013; Liu et al., 2020). The present study is not a clinical trial, so these guidelines are used as structural inspiration rather than as formal regulatory requirements.

+


+

## 2. Study Objectives

+


+

### 2.1 Primary Objective

+


+

To determine whether use of a conversational AI mediator improves team-level transparency, as measured by change in Team Transparency Index (TTI) from baseline to intervention.

+


+

### 2.2 Secondary Objectives

+


+

1. To estimate whether TTI can be coded with acceptable inter-rater reliability in real Agile work artifacts.

+

2. To estimate the technical performance of AI extraction, artifact linking, and conflict detection against a manually coded gold sample.

+

3. To determine whether the intervention reduces mean time to detect communication-to-record inconsistencies.

+

4. To determine whether the intervention improves documentation completeness for action items and decisions.

+

5. To assess whether the intervention changes perceived transparency, cognitive workload, trust in AI-generated updates, and psychological safety.

+

6. To characterize qualitative adoption patterns, including when teams accept, correct, ignore, or reject AI-generated prompts.

+

7. To identify governance risks associated with AI-mediated team memory, especially surveillance concerns and misuse of transparency metrics for individual evaluation.

+


+

## 3. Research Questions and Hypotheses

+


+

**RQ1:** Does the conversational AI mediator improve team transparency compared with baseline practice?

+


+

**H1a:** In the feasibility pilot, TTI components will reach acceptable coding reliability, defined as Cohen's kappa or Krippendorff's alpha >= 0.70 for categorical judgments and intraclass correlation >= 0.70 for continuous timing measures.

+


+

**H1b:** In the powered field evaluation, mean TTI will be higher during the intervention period than during the baseline period.

+


+

**RQ2:** Does the mediator improve detection of mismatches between team communication and project records?

+


+

**H2:** Mean time to detect communication-to-record inconsistencies will be lower during intervention than during baseline.

+


+

**RQ3:** Does the mediator improve documentation quality without increasing perceived burden?

+


+

**H3a:** Documentation completeness will increase during intervention.

+


+

**H3b:** Perceived workload will not increase by more than 10 points on a 0-100 raw NASA-TLX scale, and AI prompts will not exceed a median of two prompts per participant per workday.

+


+

**RQ4:** How do team members experience AI-mediated confirmation prompts?

+


+

**H4:** Team members will report higher trust in AI-generated updates when prompts include source links, confidence levels, and reversible writeback.

+


+

**RQ5:** What governance safeguards are necessary to preserve psychological safety?

+


+

This question is exploratory and will be answered through interviews, observations, and thematic analysis.

+


+

## 4. Study Design

+


+

The study will use a two-stage mixed-method design.

+


+

**Stage A: Feasibility pilot.** Stage A tests whether the protocol can be implemented safely and whether TTI can be measured reliably. It estimates coding reliability, AI technical performance, prompt burden, missingness, opt-out rates, and safety signals. Stage A is not powered to test effectiveness.

+


+

**Stage B: Field evaluation.** Stage B proceeds only if Stage A meets progression criteria. It uses a quasi-experimental repeated-measures design in which each participating team completes a baseline phase followed by an intervention phase. If organizational scheduling permits, teams will be staggered so that not all teams begin the intervention simultaneously. Staggering improves interpretability by separating intervention effects from calendar events such as release deadlines or organizational changes.

+


+

The recommended minimum duration is:

+


+

| Phase | Duration | Purpose |

+

| --- | --- | --- |

+

| Preparation | 2 weeks | Consent, tool configuration, privacy review, pilot data mapping |

+

| Stage A baseline | 1 sprint or 2 weeks | Test passive data capture and initial TTI coding |

+

| Stage A intervention | 1 sprint or 2 weeks | Test prompts, technical accuracy, safety gates, and burden |

+

| Stage A decision | 1 week | Apply progression criteria before Stage B |

+

| Stage B baseline | 2 sprints or 4 weeks | Measure normal workflow without AI writeback |

+

| Stage B intervention | 2 to 4 sprints or 4 to 8 weeks | Deploy AI mediator with confirmation prompts |

+

| Follow-up | 1 to 2 weeks | Interviews, debrief, data-quality checks |

+


+

The design is not blinded. Participants will know when the AI mediator is active. Outcome extraction from system logs should be automated where possible and reviewed using predefined rules to reduce subjective bias.

+


+

## 5. Setting

+


+

The study will be conducted in software development teams that use Agile practices and maintain digital project artifacts. The minimum tooling environment is:

+


+

1. An issue tracker such as Jira.

+

2. A Git-based version-control system.

+

3. A team communication channel such as Rocket.Chat, Slack, Microsoft Teams, or equivalent.

+

4. Recurring stand-up communication, either synchronous or asynchronous.

+


+

The initial pilot may use a single Scrum team of 6 to 10 members. A stronger field evaluation should include at least 6 teams to support team-level comparison and reduce the risk that findings reflect one team's habits. Baseline tooling must be documented for each team, including existing bots, meeting summarizers, Jira automation rules, issue-linking practices, and dashboard use.

+


+

## 6. Participants

+


+

### 6.1 Target Population

+


+

Participants are members of Agile software development teams, including developers, QA engineers, product owners, scrum masters, engineering managers, and other roles who participate in daily coordination or issue updates.

+


+

### 6.2 Inclusion Criteria

+


+

1. The participant is a member of a participating Agile team.

+

2. The participant uses the team's issue tracker, code review system, or communication channel as part of normal work.

+

3. The participant is at least 18 years old.

+

4. The participant provides informed consent for study data collection.

+


+

### 6.3 Exclusion Criteria

+


+

1. Participants who do not consent to data collection.

+

2. Contractors or external stakeholders whose communication cannot be captured under the organization's data policy.

+

3. Team members whose role creates a direct power conflict that cannot be mitigated in consent or interview procedures.

+


+

### 6.4 Sampling Strategy

+


+

Team recruitment will use purposive sampling. The study should prioritize teams with active project work, regular stand-up practices, and enough tool usage to support traceability measurement. Within recruited teams, all eligible members should be invited to participate to reduce selection bias.

+


+

### 6.5 Target Sample Size

+


+

For Stage A, the target is 1 to 3 teams and approximately 8 to 30 participants. Stage A will be judged by feasibility rather than statistical significance. Progression to Stage B requires: (a) TTI coding reliability >= 0.70, (b) median prompt burden <= 2 prompts per participant per workday, (c) no unresolved ethics or safety gate breach, (d) AI extraction/linking precision >= 0.70 on the manually coded pilot sample, and (e) no more than 20% missingness in the primary data streams.

+


+

For Stage B, the target should be at least 6 teams if feasible. A final power analysis will be completed after Stage A using the observed TTI variance, estimated intraclass correlation, team count, sprint count, and expected missingness. If the available team count is too small for confirmatory inference, Stage B will be reported as an expanded feasibility and estimation study rather than an effectiveness trial.

+


+

## 7. Intervention

+


+

### 7.1 Conversational AI Mediator

+


+

The intervention is a conversational AI framework embedded in the team's communication and project-management environment. It performs four functions:

+


+

1. **Ingestion:** Collects meeting transcripts, chat messages, Jira updates, and Git metadata.

+

2. **Extraction and linking:** Identifies candidate action items, decisions, blockers, status claims, and links to project artifacts.

+

3. **Conflict detection:** Flags mismatches between communication and project records.

+

4. **Role-aware confirmation:** Prompts relevant team members to confirm, reject, or edit candidate knowledge artifacts before writeback.

+


+

The study will record the exact model family, version, system prompts, retrieval configuration, source connectors, and confidence-scoring rules used during the intervention. Any model or prompt change during data collection will be logged as a protocol deviation and sensitivity-analysis flag.

+


+

### 7.2 Prompt Taxonomy

+


+

The mediator may generate five prompt types:

+


+

| Prompt Type | Trigger | Required Recipient |

+

| --- | --- | --- |

+

| Action-item confirmation | Candidate who/what/when commitment extracted from communication | Named assignee |

+

| Decision confirmation | Scope, priority, acceptance criterion, or architecture decision detected | Product owner plus affected implementer |

+

| Blocker clarification | Blocker stated without owner, dependency, or next step | Blocked assignee and blocker owner when identifiable |

+

| Conflict resolution | Communication claim conflicts with Jira/Git status | Assignee plus relevant role based on artifact type |

+

| Documentation completion | Confirmed item lacks required metadata | Assignee or scrum master/team lead |

+


+

Prompt content must include the extracted claim, source link, linked artifact, confidence level, suggested action, and available responses: accept, edit, reject, defer, or mark sensitive.

+


+

### 7.3 Confidence and Escalation Rules

+


+

The mediator will use predefined confidence bands. High-confidence, low-impact items may be batched. Medium-confidence items require explicit confirmation. Low-confidence items are logged for technical evaluation but do not trigger participant prompts unless sampled for manual review. High-impact items always require confirmation regardless of confidence.

+


+

| Confidence Band | Operational Rule |

+

| --- | --- |

+

| High | Confidence >= 0.80 and no conflict detected; batch unless high-impact. |

+

| Medium | 0.50 <= confidence < 0.80; request confirmation before writeback. |

+

| Low | Confidence < 0.50; do not prompt by default; include in manual evaluation sample. |

+

| High-impact override | Scope, priority, ownership, due date, acceptance criteria, security/privacy, or release decision; require role-aware confirmation. |

+


+

### 7.4 Technical Performance Evaluation

+


+

A stratified random sample of communication events and AI-generated candidates will be manually coded by two independent coders. The gold sample will include accepted prompts, edited prompts, rejected prompts, ignored prompts, and low-confidence non-prompted candidates. Technical outcomes will include precision, recall where denominators can be estimated, F1 score, false-positive categories, false-negative categories, and disagreement resolution notes for extraction, artifact linking, and conflict detection.

+


+

### 7.5 Human Oversight

+


+

The AI does not independently decide project scope, task status, ownership, or acceptance criteria. For low-risk summaries, one assignee confirmation may be sufficient. For high-impact updates, confirmation may be required from multiple roles, such as developer, QA, and product owner.

+


+

### 7.6 Writeback Policy

+


+

Confirmed updates may be written to Jira comments, issue metadata, pull request descriptions, decision records, or a team knowledge store. Writeback must preserve provenance. Every AI-created record should include source links, timestamp, confirming roles, and reversal instructions.

+


+

The mediator may write comments or draft suggestions automatically after confirmation, but it may not silently change issue status, assignee, due date, sprint scope, acceptance criteria, or release labels. Those fields require explicit role-aware confirmation and a reversible audit trail.

+


+

### 7.7 Prompt Governance and Participant Controls

+


+

Prompt frequency will be capped to reduce interruption burden. The default cap is two prompts per participant per workday, excluding urgent high-impact conflicts. Participants may pause prompts for a defined period, mark a source as sensitive, reject a prompt without justification, edit the proposed record, request deletion from the study dataset when allowed by policy, or appeal a writeback to the data steward. The system should batch low-risk suggestions and prioritize prompts involving conflict, missing ownership, missing due date, unresolved blocker, or high-impact decision.

+


+

## 8. Comparator

+


+

The comparator is each team's baseline workflow without AI-mediated extraction, confirmation, or writeback. Teams continue to use their existing communication channels, issue trackers, stand-ups, and version-control practices.

+


+

## 9. Outcomes

+


+

### 9.1 Primary Outcome

+


+

The primary outcome is change in Team Transparency Index (TTI) from baseline to intervention.

+


+

```text

+

TTI = 0.25*COV + 0.25*CON + 0.20*CSN + 0.15*TML + 0.15*CMP

+

```

+


+

These weights are theory-informed a priori weights and will be fixed before data collection. Any change to the weights, component definitions, or component inclusion rules will require a documented protocol amendment and will not be tuned on outcome data.

+


+

| Component | Operational Definition |

+

| --- | --- |

+

| COV | Proportion of eligible communication-mentioned tasks or decisions linked to a Jira issue, Git artifact, or decision record within the sprint window. Denominator excludes social talk, duplicate mentions, and explicitly out-of-scope personal content. |

+

| CON | Proportion of sampled status, ownership, blocker, and decision claims that match structured project records or are explicitly reconciled. |

+

| CSN | Proportion of high-impact decisions that meet the predefined role-confirmation threshold before writeback. |

+

| TML | Normalized inverse delay between event occurrence and documented update, capped at the sprint boundary. |

+

| CMP | Proportion of eligible action items with who, what, and when fields. "When" may be a due date, sprint, next meeting, or explicit "no date yet" confirmation. |

+


+

TTI will be coded at sprint level for each team. The sampling window is the sprint plus a 48-hour post-sprint reconciliation period for records that are updated immediately after review or retrospective discussion. Eligible communication events are stand-up statements, chat messages, issue comments, pull request comments, and meeting transcript segments that contain a task, blocker, status claim, ownership claim, decision, due-date claim, or acceptance-criteria claim. Events are excluded when they are duplicate reminders, social conversation, private personnel content, or communication from non-consenting participants that cannot be de-identified under the approved protocol.

+


+

The denominator for each component is constructed independently. For example, a statement such as "Ana will add rate-limit handling before release candidate 2" contributes to COV if it can be linked to an issue or pull request, to CON if the claim matches project records or is explicitly reconciled, to TML based on the time until the linked record is updated, and to CMP if the responsible person, work item, and timing are present. A scope decision such as "we are deferring SSO to the next sprint" contributes to CSN if it meets the predefined product-owner and affected-implementer confirmation rule.

+


+

Two coders will independently code a 20% stratified sample of events during Stage A, covering each data source, role, prompt type, and confidence band. Disagreements will be adjudicated by a third reviewer. The study will report component-level reliability and will freeze the coding manual before Stage B only if the reliability progression criterion is met.

+


+

### 9.2 Secondary Outcomes

+


+

| Outcome | Measurement |

+

| --- | --- |

+

| Mean time to detect inconsistency | Time from first conflicting signal to system or human identification. |

+

| Documentation completeness | Share of action items and decisions with complete metadata. |

+

| Prompt burden | Number of AI prompts per participant per workday and participant-rated interruption cost. |

+

| Trust in AI updates | Survey items assessing perceived accuracy, explainability, and control. |

+

| Workload | NASA-TLX or raw NASA-TLX adapted for subjective workload measurement (Hart, 2006). |

+

| Psychological safety | Team psychological safety survey based on Edmondson's construct (Edmondson, 1999). |

+

| Adoption behavior | Acceptance, edit, rejection, and ignore rates for AI suggestions. |

+

| Technical extraction accuracy | Precision, recall where estimable, and F1 for action/decision/blocker extraction. |

+

| Technical linking accuracy | Accuracy and false-link rate for Jira/Git/decision-record links. |

+

| Conflict-detection accuracy | Precision and false-negative categories against manually coded conflict samples. |

+

| Governance concerns | Interview-coded concerns about privacy, surveillance, accountability, and misuse. |

+


+

## 10. Variables

+


+

| Role | Variable | Measurement | Scale |

+

| --- | --- | --- | --- |

+

| Intervention | AI mediator active | Baseline = 0, intervention = 1 | Nominal |

+

| Primary DV | TTI | Weighted composite of five normalized components | Interval |

+

| Secondary DV | Inconsistency detection time | Hours from conflict creation to detection | Ratio |

+

| Secondary DV | Documentation completeness | Complete items / total action items | Ratio |

+

| Secondary DV | Prompt burden | Prompts per user per day; survey burden rating | Ratio / ordinal |

+

| Secondary DV | Workload | NASA-TLX or raw NASA-TLX score | Interval |

+

| Secondary DV | Psychological safety | Mean survey score | Interval |

+

| Control | Team size | Number of active team members | Ratio |

+

| Control | Sprint phase | Planning, execution, release, retrospective | Nominal |

+

| Control | Workload intensity | Issue count, pull request count, release deadline indicator | Ratio / nominal |

+

| Confound | Management pressure | Interview and survey indicators | Qualitative / ordinal |

+

| Confound | Tool maturity | Baseline completeness and issue hygiene | Interval |

+


+

## 11. Instruments and Data Sources

+


+

| Instrument or Source | Purpose | Notes |

+

| --- | --- | --- |

+

| Jira or issue tracker export | Issue status, assignee, transitions, comments, timestamps | Metadata and project records only unless consent allows content review. |

+

| Git hosting metadata | Commits, pull requests, reviews, branch references | Used for traceability linking and event timing. |

+

| Chat and stand-up transcripts | Candidate decisions, blockers, status claims, action items | Sensitive content redaction required. |

+

| AI prompt log | Prompt type, recipient role, source evidence, confidence band, response, confirmation outcome, override reason | Used for adoption, burden, and safety-gate analysis. |

+

| TTI extraction rubric | Standardized coding of COV, CON, CSN, TML, CMP | Frozen after Stage A if reliability criteria are met. |

+

| Technical gold sample | Manually coded sample of candidate action items, decisions, blockers, links, and conflicts | Used to estimate extraction, linking, and conflict-detection performance. |

+

| Perceived transparency survey | Participant perception of alignment and visibility | Administer baseline and intervention. |

+

| Trust in AI update survey | Explainability, source confidence, reversibility, perceived accuracy | Administer after intervention. |

+

| NASA-TLX or raw NASA-TLX | Subjective workload | Use consistently across phases. |

+

| Psychological safety survey | Team climate for interpersonal risk taking | Use validated or adapted items with permission where required. |

+

| Semi-structured interviews | Qualitative adoption and governance data | Conduct after intervention. |

+


+

## 12. Data Collection Procedure

+


+

### 12.1 Preparation

+


+

The research team will obtain organizational approval, ethics approval where required, and informed consent. Tool integrations will be configured with least-privilege access. A data mapping exercise will identify which fields are needed for outcome measurement and which fields must be excluded or redacted.

+


+

### 12.2 Baseline Phase

+


+

During baseline, the AI mediator will not prompt participants or write back to project systems. Data will be collected passively from agreed sources to compute baseline TTI and secondary measures. Participants will complete baseline surveys on perceived transparency, workload, and psychological safety.

+


+

### 12.3 Intervention Phase

+


+

During intervention, the AI mediator will generate candidate knowledge items and role-aware prompts. Participants may accept, edit, reject, or ignore prompts. Confirmed items may be written back according to the writeback policy. System logs will record prompt type, source evidence, confirmation route, and outcome.

+


+

### 12.4 Follow-up

+


+

Participants will complete post-intervention surveys. A purposive subset of participants across roles will be invited for interviews. Interviews will focus on usefulness, trust, interruptions, missed cases, false positives, correction behavior, and privacy concerns.

+


+

## 13. Analysis Plan

+


+

### 13.1 Stage A Feasibility Analysis

+


+

Stage A will be analyzed as a feasibility pilot. The primary outputs will be recruitment and retention rates, consent coverage by data source, missingness by variable, prompt burden, safety-gate events, TTI coding reliability, and technical performance against the manually coded gold sample. Progression to Stage B requires meeting the criteria in Section 6. If criteria are not met, the study will be reported as a feasibility study and the intervention or protocol will be revised before any confirmatory evaluation.

+


+

### 13.2 Quantitative Effectiveness Analysis

+


+

If Stage B proceeds, the primary analysis will compare TTI between baseline and intervention. If multiple teams are included, mixed-effects models should be used with phase as a fixed effect and team as a random effect. If the sample is too small or assumptions are not met, the analysis will report descriptive statistics, paired comparisons, effect sizes, and confidence intervals.

+


+

Secondary outcomes will be analyzed as follows:

+


+

| Outcome | Primary Test | Fallback |

+

| --- | --- | --- |

+

| TTI | Linear mixed-effects model | Wilcoxon signed-rank or descriptive effect size |

+

| Detection time | Survival or time-to-event model | Mann-Whitney U or paired non-parametric comparison |

+

| Documentation completeness | Logistic or beta regression | Proportion difference with confidence interval |

+

| Prompt burden | Poisson or negative binomial model | Descriptive rate comparison |

+

| Survey scales | Paired t-test or mixed model | Wilcoxon signed-rank |

+

| Adoption behavior | Acceptance/edit/rejection rates | Descriptive and role-stratified analysis |

+

| Technical performance | Precision, recall, F1, false-positive and false-negative review | Descriptive error taxonomy |

+


+

Multiple comparisons will be treated as exploratory unless the study is powered confirmatorily. Survey scale reliability will be assessed with internal consistency when sample size permits.

+


+

### 13.3 Missing Data and Partial Consent

+


+

Missing data will be reported by variable, phase, team, role, and source system. Partial consent will be handled by excluding non-consenting participants' identifiable content from qualitative analysis and by using only aggregate or de-identified metadata where permitted by the approved consent protocol. If consent gaps prevent reliable team-level TTI computation, that team-period will be excluded from primary effectiveness analysis and retained only for feasibility reporting. Sensitivity analyses will compare complete-case results with analyses using available de-identified metadata where ethically and statistically appropriate.

+


+

### 13.4 Qualitative Analysis

+


+

Interview transcripts and observation notes will be analyzed using thematic analysis. The initial coding frame will include trust, usefulness, interruption cost, correction behavior, false positives, missed updates, role conflict, surveillance concern, and psychological safety. Two coders should independently code a subset of material and reconcile disagreements before coding the full dataset.

+


+

### 13.5 Mixed-Method Integration

+


+

Quantitative and qualitative findings will be integrated through joint displays. For example, teams with increased TTI but high prompt burden will be examined qualitatively to determine whether the transparency gain was acceptable. Teams with low AI adoption will be examined for trust, workflow fit, and governance barriers.

+


+

## 14. Data Management

+


+

Data will be minimized to the fields required for the study. Raw chat or transcript content should be redacted or summarized when possible. Identifiers will be pseudonymized before analysis. A linkage file, if required, will be stored separately with restricted access. Data storage will use encrypted institutional or organizational storage. Retention period should be defined before data collection and communicated in the consent form.

+


+

The TTI should be reported at team level. Individual-level prompt response data may be needed for analysis, but it must not be used for performance evaluation. Any publication should aggregate or anonymize examples to prevent re-identification.

+


+

Data access will be separated by role. The principal investigator may access the full approved research dataset; the data steward may access linkage and redaction files; the technical integration owner may access system logs needed for debugging but not interview material; and team representatives may review only aggregated disclosure summaries. Managers will receive team-level summaries only after aggregation and disclosure review.

+


+

## 15. Ethics and Governance

+


+

This study involves workplace communication and therefore creates privacy and power-differential risks. The following safeguards are required:

+


+

1. Participation must be voluntary and based on informed consent.

+

2. Team members must know which channels and artifacts are included.

+

3. Managers must not receive individual-level transparency or prompt-response scores.

+

4. AI-generated updates must be visible, source-linked, and reversible.

+

5. Sensitive personal content must be excluded or redacted.

+

6. Participants must be able to challenge or correct AI-generated records.

+

7. Interview participation must be separated from management evaluation.

+

8. Participants must be able to pause prompts, mark content sensitive, reject or edit suggested updates, request deletion of erroneous AI-generated records, and appeal contested records through a named governance route.

+

9. Manager-facing dashboards must exclude individual prompt-response behavior, individual transparency scores, and any ranking or comparison of named participants.

+


+

Psychological safety is both an outcome and an ethical constraint. Edmondson (1999) defines psychological safety as a shared belief that the team is safe for interpersonal risk taking. A transparency tool that makes people afraid to surface blockers would fail even if it improves documentation metrics.

+


+

### 15.1 Safety Gates and Stopping Rules

+


+

The intervention will be paused for governance review if any of the following occur:

+


+

1. Mean psychological safety drops by 0.5 or more on a five-point scale from baseline.

+

2. Mean raw NASA-TLX workload increases by more than 10 points from baseline.

+

3. Median prompt burden exceeds two prompts per participant per workday for two consecutive weeks.

+

4. More than 20% of participants use pause, opt-out, or mark-sensitive controls during a sprint.

+

5. Any participant reports perceived retaliation, coercion, or manager misuse linked to AI-mediated records.

+

6. Sensitive data are captured outside the approved source scope and are not remediated within the incident response window.

+


+

TTI improvement will be considered acceptable only if these safety thresholds remain within bounds. A team that improves TTI while breaching safety thresholds will be reported as a governance failure rather than an effectiveness success.

+


+

## 16. Risk Management

+


+

| Risk | Likelihood | Impact | Mitigation |

+

| --- | --- | --- | --- |

+

| Excessive prompts interrupt work | Medium | Medium | Prompt caps, batching, priority rules, pause controls, safety-gate review. |

+

| False positives reduce trust | Medium | Medium | Source links, confidence labels, easy rejection/editing. |

+

| Surveillance perception | Medium | High | Consent, team-level reporting, no individual scoring, governance review. |

+

| Sensitive data captured | Medium | High | Channel scoping, redaction, access controls, data minimization. |

+

| Tool integration failure | Medium | Medium | Pilot mapping, fallback export, manual coding sample. |

+

| Management misuse | Low to medium | High | Written policy forbidding individual performance use. |

+

| AI writeback error | Medium | Medium | Human confirmation, reversible updates, audit trail, deletion request and appeal route. |

+


+

## 17. Dissemination Plan

+


+

Findings will be disseminated as a protocol paper, a field evaluation paper after data collection, and a practitioner-oriented report for participating teams. The field evaluation report will separate confirmed findings from exploratory observations and will disclose limitations related to sample size, team context, and tool configuration.

+


+

Before recruitment, the study team will select a preregistration destination and artifact repository, such as OSF or an institutional repository, and will specify which protocol materials, de-identified analysis code, instrument templates, and non-sensitive aggregate outputs can be shared. The final field evaluation report will include a reporting checklist adapted from protocol-reporting and AI-intervention reporting guidance.

+


+

## 18. Protocol Status

+


+

This is a revised protocol draft prepared for preregistration and ethics review. It should not be treated as preregistered, ethics-approved, or implementation-ready until the following items are completed:

+


+

1. Participating organization and teams.

+

2. Exact tool integrations and data fields.

+

3. Consent language.

+

4. Survey instruments and permissions.

+

5. Prompt governance thresholds.

+

6. Statistical analysis plan details based on expected team count and sprint duration.

+


+

## 19. Appendix Roadmap

+


+

The final preregistration package should include the following study artifacts:

+


+

1. Consent form and participant information sheet.

+

2. Data source map and data dictionary.

+

3. TTI coding manual and adjudication guide.

+

4. Prompt taxonomy and prompt examples.

+

5. Survey items for perceived transparency, trust in AI updates, workload, and psychological safety.

+

6. Semi-structured interview guide.

+

7. Technical gold-sample coding guide.

+

8. Safety incident form and escalation workflow.

+

9. Statistical analysis plan and analysis code plan.

+


+

## References

+


+

Ball, K. (2021). *Electronic monitoring and surveillance in the workplace: Literature review and policy recommendations*. Publications Office of the European Union. https://doi.org/10.2760/451453

+


+

Chan, A.-W., Tetzlaff, J. M., Altman, D. G., Laupacis, A., Gotzsche, P. C., Krleza-Jeric, K., Hrobjartsson, A., Mann, H., Dickersin, K., Berlin, J. A., Dore, C. J., Parulekar, W. R., Summerskill, W. S. M., Groves, T., Schulz, K. F., Sox, H. C., Rockhold, F. W., Rennie, D., & Moher, D. (2013). SPIRIT 2013 statement: Defining standard protocol items for clinical trials. *Annals of Internal Medicine, 158*(3), 200-207. https://doi.org/10.7326/0003-4819-158-3-201302050-00583

+


+

Cinkusz, K., Chudziak, J. A., & Niewiadomska-Szynkiewicz, E. (2025). Cognitive agents powered by large language models for agile software project management. *Electronics, 14*(1), Article 87. https://doi.org/10.3390/electronics14010087

+


+

Cleland-Huang, J., Gotel, O. C. Z., Huffman Hayes, J., Mäder, P., & Zisman, A. (2014). Software traceability: Trends and future directions. *Future of Software Engineering Proceedings*, 55-69. https://doi.org/10.1145/2593882.2593891

+


+

Edmondson, A. (1999). Psychological safety and learning behavior in work teams. *Administrative Science Quarterly, 44*(2), 350-383. https://doi.org/10.2307/2666999

+


+

Hart, S. G. (2006). NASA-Task Load Index (NASA-TLX); 20 years later. *Proceedings of the Human Factors and Ergonomics Society Annual Meeting, 50*(9), 904-908. https://doi.org/10.1177/154193120605000909

+


+

Itzik, D., & Roy, G. (2023). Does agile methodology fit all characteristics of software projects? Review and analysis. *Empirical Software Engineering, 28*, Article 105. https://doi.org/10.1007/s10664-023-10334-7

+


+

Liu, X., Cruz Rivera, S., Moher, D., Calvert, M. J., Denniston, A. K., & SPIRIT-AI and CONSORT-AI Working Group. (2020). Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: The CONSORT-AI extension. *Nature Medicine, 26*, 1364-1374. https://doi.org/10.1038/s41591-020-1034-x

+


+

Malla, P. (2025). Analyzing the impact of agile methodologies on software quality and delivery speed: A comparative study. *World Journal of Advanced Research and Reviews, 25*(1), 1207-1216. https://doi.org/10.30574/wjarr.2025.25.1.0184

+


+

Parasuraman, R., & Riley, V. (1997). Humans and automation: Use, misuse, disuse, abuse. *Human Factors, 39*(2), 230-253. https://doi.org/10.1518/001872097778543886

+


+

Shneiderman, B. (2020). Human-centered artificial intelligence: Reliable, safe & trustworthy. *International Journal of Human-Computer Interaction, 36*(6), 495-504. https://doi.org/10.1080/10447318.2020.1741118

+


+

Stray, V., Moe, N. B., & Bergersen, G. R. (2017). Are daily stand-up meetings valuable? A survey of developers in software teams. In H. Baumeister, H. Lichter, & M. Riebisch (Eds.), *Agile Processes in Software Engineering and Extreme Programming* (pp. 274-281). Springer. https://doi.org/10.1007/978-3-319-57633-6_20

+


+

Stray, V., Moe, N. B., & Sjoberg, D. I. K. (2020). Daily stand-up meetings: Start breaking the rules. *IEEE Software, 37*(3), 70-77. https://doi.org/10.1109/MS.2018.2875988

+


+

Stray, V., Sjoberg, D. I. K., & Dyba, T. (2016). The daily stand-up meeting: A grounded theory study. *Journal of Systems and Software, 114*, 101-124. https://doi.org/10.1016/j.jss.2016.01.004

+


+

Umar, M. A. M. A., Lano, K., & Abubakar, A. K. (2025). Automated requirements engineering framework for agile model-driven development. *Frontiers in Computer Science, 7*, Article 1537100. https://doi.org/10.3389/fcomp.2025.1537100

+


+

Verwijs, C., & Russo, D. (2024). Do Agile scaling approaches make a difference? An empirical comparison of team effectiveness across popular scaling approaches. *Empirical Software Engineering, 29*, Article 75. https://doi.org/10.1007/s10664-024-10481-5

+ + diff --git a/Aidaily_final_manuscript.md b/Aidaily_final_manuscript.md new file mode 100644 index 0000000..a7b409c --- /dev/null +++ b/Aidaily_final_manuscript.md @@ -0,0 +1,434 @@ +# Protocol for Evaluating a Conversational AI Framework for Agile Team Transparency and Knowledge Traceability + +## Abstract + +**Background:** Agile teams rely on daily coordination, issue tracking, and version control to maintain shared understanding. In distributed and hybrid teams, however, decisions, blockers, and action commitments often remain fragmented across meetings, chat, Jira, and Git. Conversational AI may help by extracting candidate updates from informal communication, linking them to project artifacts, and prompting team members to confirm or correct them. + +**Objective:** This protocol describes a mixed-method field study to evaluate whether a conversational AI framework improves Agile team transparency, knowledge traceability, and perceived alignment without increasing cognitive burden or weakening psychological safety. + +**Methods:** The study will use a two-stage mixed-method design. Stage A is a feasibility pilot that estimates measurement reliability, prompt burden, extraction/linking accuracy, and safety signals. Stage B is an optional powered field evaluation using a quasi-experimental, baseline-to-intervention design across Agile software teams. During baseline, teams continue normal workflows while communication, issue, and version-control metadata are measured. During intervention, teams use a conversational AI mediator that ingests stand-up transcripts, chat messages, Jira data, and Git metadata; extracts candidate decisions and action items; detects inconsistencies; and requests role-aware confirmation before writeback. The primary feasibility outcome is whether the Team Transparency Index (TTI) can be computed reliably. The primary effectiveness outcome, if Stage B proceeds, is change in TTI from baseline to intervention. + +**Analysis:** Feasibility analysis will report reliability, prompt-burden rates, technical performance, missingness, and progression criteria. Effectiveness analysis will compare baseline and intervention periods using mixed-effects models or non-parametric alternatives when assumptions are not met. Qualitative interviews and observation notes will be analyzed thematically to explain adoption patterns, trust, interruption costs, and governance concerns. + +**Ethics and Dissemination:** The study requires informed consent, role-based access controls, provenance-preserving audit trails, and safeguards against individual performance scoring. Results will be reported as a protocol-compliant field evaluation and will distinguish system performance from team-level organizational outcomes. + +**Keywords:** Agile software development; protocol paper; conversational AI; mixed methods; field study; team transparency; knowledge traceability; psychological safety; human-AI collaboration + +## 1. Introduction + +Agile software development depends on shared context. Teams coordinate through daily stand-ups, chat threads, issue trackers, code reviews, and version-control activity. These artifacts are individually useful, but they do not automatically produce a durable and verified team memory. A decision may be made verbally, partly clarified in chat, reflected indirectly in a pull request, and never updated in the issue tracker. The result is a persistent gap between what the team knows informally and what the system of record says formally. + +Prior research on daily stand-ups shows both the value and limits of recurring Agile communication. Stand-ups are widely used and can support team awareness, but their value varies by team size, role, and meeting quality (Stray et al., 2017). Grounded theory work also shows that stand-ups support coordination and monitoring while remaining sensitive to local practice (Stray et al., 2016). Later work argues that teams should adapt stand-up rules when the ritual no longer serves communication needs (Stray et al., 2020). These findings suggest that Agile transparency cannot be inferred from ceremony adoption alone. + +Research on Agile scaling and methodology fit also supports a context-sensitive approach. Verwijs and Russo (2024) found that scaling frameworks themselves explain little practical difference in team effectiveness. Itzik and Roy (2023) argue that Agile fit depends on software project characteristics and should be assessed through a decision framework. For a transparency intervention, this means the target is not framework compliance but the quality of alignment among people, artifacts, and decisions. + +AI-supported software project management provides relevant technical foundations. Automated requirements engineering research shows that machine learning can extract structured models from natural-language requirements in Agile contexts (Umar et al., 2025). LLM-based multi-agent project-management frameworks such as CogniSim show that AI agents can support Agile roles and project workflows in simulated environments (Cinkusz et al., 2025). Comparative work on Agile methods and technology-enhanced practices also suggests that AI-enabled tools may affect delivery and quality outcomes, while introducing risks of over-reliance on automation (Malla, 2025). The traceability literature is also directly relevant: software traceability is valuable but often performed ad hoc and after the fact, limiting its realized benefit (Cleland-Huang et al., 2014). Human-centered AI research further emphasizes that reliable and trustworthy systems should combine high automation with meaningful human control (Shneiderman, 2020). These strands motivate a field study that tests whether conversational AI can improve the human communication layer of real Agile teams without producing automation bias, intrusive monitoring, or reduced psychological safety (Parasuraman & Riley, 1997; Ball, 2021). + +This protocol defines a study to evaluate a conversational AI mediator for Agile transparency. The protocol is informed by general protocol-reporting principles for transparent intervention studies, including the logic of SPIRIT-style completeness and AI-specific reporting attention to intervention behavior, human oversight, and error handling (Chan et al., 2013; Liu et al., 2020). The present study is not a clinical trial, so these guidelines are used as structural inspiration rather than as formal regulatory requirements. + +## 2. Study Objectives + +### 2.1 Primary Objective + +To determine whether use of a conversational AI mediator improves team-level transparency, as measured by change in Team Transparency Index (TTI) from baseline to intervention. + +### 2.2 Secondary Objectives + +1. To estimate whether TTI can be coded with acceptable inter-rater reliability in real Agile work artifacts. +2. To estimate the technical performance of AI extraction, artifact linking, and conflict detection against a manually coded gold sample. +3. To determine whether the intervention reduces mean time to detect communication-to-record inconsistencies. +4. To determine whether the intervention improves documentation completeness for action items and decisions. +5. To assess whether the intervention changes perceived transparency, cognitive workload, trust in AI-generated updates, and psychological safety. +6. To characterize qualitative adoption patterns, including when teams accept, correct, ignore, or reject AI-generated prompts. +7. To identify governance risks associated with AI-mediated team memory, especially surveillance concerns and misuse of transparency metrics for individual evaluation. + +## 3. Research Questions and Hypotheses + +**RQ1:** Does the conversational AI mediator improve team transparency compared with baseline practice? + +**H1a:** In the feasibility pilot, TTI components will reach acceptable coding reliability, defined as Cohen's kappa or Krippendorff's alpha >= 0.70 for categorical judgments and intraclass correlation >= 0.70 for continuous timing measures. + +**H1b:** In the powered field evaluation, mean TTI will be higher during the intervention period than during the baseline period. + +**RQ2:** Does the mediator improve detection of mismatches between team communication and project records? + +**H2:** Mean time to detect communication-to-record inconsistencies will be lower during intervention than during baseline. + +**RQ3:** Does the mediator improve documentation quality without increasing perceived burden? + +**H3a:** Documentation completeness will increase during intervention. + +**H3b:** Perceived workload will not increase by more than 10 points on a 0-100 raw NASA-TLX scale, and AI prompts will not exceed a median of two prompts per participant per workday. + +**RQ4:** How do team members experience AI-mediated confirmation prompts? + +**H4:** Team members will report higher trust in AI-generated updates when prompts include source links, confidence levels, and reversible writeback. + +**RQ5:** What governance safeguards are necessary to preserve psychological safety? + +This question is exploratory and will be answered through interviews, observations, and thematic analysis. + +## 4. Study Design + +The study will use a two-stage mixed-method design. + +**Stage A: Feasibility pilot.** Stage A tests whether the protocol can be implemented safely and whether TTI can be measured reliably. It estimates coding reliability, AI technical performance, prompt burden, missingness, opt-out rates, and safety signals. Stage A is not powered to test effectiveness. + +**Stage B: Field evaluation.** Stage B proceeds only if Stage A meets progression criteria. It uses a quasi-experimental repeated-measures design in which each participating team completes a baseline phase followed by an intervention phase. If organizational scheduling permits, teams will be staggered so that not all teams begin the intervention simultaneously. Staggering improves interpretability by separating intervention effects from calendar events such as release deadlines or organizational changes. + +The recommended minimum duration is: + +| Phase | Duration | Purpose | +| --- | --- | --- | +| Preparation | 2 weeks | Consent, tool configuration, privacy review, pilot data mapping | +| Stage A baseline | 1 sprint or 2 weeks | Test passive data capture and initial TTI coding | +| Stage A intervention | 1 sprint or 2 weeks | Test prompts, technical accuracy, safety gates, and burden | +| Stage A decision | 1 week | Apply progression criteria before Stage B | +| Stage B baseline | 2 sprints or 4 weeks | Measure normal workflow without AI writeback | +| Stage B intervention | 2 to 4 sprints or 4 to 8 weeks | Deploy AI mediator with confirmation prompts | +| Follow-up | 1 to 2 weeks | Interviews, debrief, data-quality checks | + +The design is not blinded. Participants will know when the AI mediator is active. Outcome extraction from system logs should be automated where possible and reviewed using predefined rules to reduce subjective bias. + +## 5. Setting + +The study will be conducted in software development teams that use Agile practices and maintain digital project artifacts. The minimum tooling environment is: + +1. An issue tracker such as Jira. +2. A Git-based version-control system. +3. A team communication channel such as Rocket.Chat, Slack, Microsoft Teams, or equivalent. +4. Recurring stand-up communication, either synchronous or asynchronous. + +The initial pilot may use a single Scrum team of 6 to 10 members. A stronger field evaluation should include at least 6 teams to support team-level comparison and reduce the risk that findings reflect one team's habits. Baseline tooling must be documented for each team, including existing bots, meeting summarizers, Jira automation rules, issue-linking practices, and dashboard use. + +## 6. Participants + +### 6.1 Target Population + +Participants are members of Agile software development teams, including developers, QA engineers, product owners, scrum masters, engineering managers, and other roles who participate in daily coordination or issue updates. + +### 6.2 Inclusion Criteria + +1. The participant is a member of a participating Agile team. +2. The participant uses the team's issue tracker, code review system, or communication channel as part of normal work. +3. The participant is at least 18 years old. +4. The participant provides informed consent for study data collection. + +### 6.3 Exclusion Criteria + +1. Participants who do not consent to data collection. +2. Contractors or external stakeholders whose communication cannot be captured under the organization's data policy. +3. Team members whose role creates a direct power conflict that cannot be mitigated in consent or interview procedures. + +### 6.4 Sampling Strategy + +Team recruitment will use purposive sampling. The study should prioritize teams with active project work, regular stand-up practices, and enough tool usage to support traceability measurement. Within recruited teams, all eligible members should be invited to participate to reduce selection bias. + +### 6.5 Target Sample Size + +For Stage A, the target is 1 to 3 teams and approximately 8 to 30 participants. Stage A will be judged by feasibility rather than statistical significance. Progression to Stage B requires: (a) TTI coding reliability >= 0.70, (b) median prompt burden <= 2 prompts per participant per workday, (c) no unresolved ethics or safety gate breach, (d) AI extraction/linking precision >= 0.70 on the manually coded pilot sample, and (e) no more than 20% missingness in the primary data streams. + +For Stage B, the target should be at least 6 teams if feasible. A final power analysis will be completed after Stage A using the observed TTI variance, estimated intraclass correlation, team count, sprint count, and expected missingness. If the available team count is too small for confirmatory inference, Stage B will be reported as an expanded feasibility and estimation study rather than an effectiveness trial. + +## 7. Intervention + +### 7.1 Conversational AI Mediator + +The intervention is a conversational AI framework embedded in the team's communication and project-management environment. It performs four functions: + +1. **Ingestion:** Collects meeting transcripts, chat messages, Jira updates, and Git metadata. +2. **Extraction and linking:** Identifies candidate action items, decisions, blockers, status claims, and links to project artifacts. +3. **Conflict detection:** Flags mismatches between communication and project records. +4. **Role-aware confirmation:** Prompts relevant team members to confirm, reject, or edit candidate knowledge artifacts before writeback. + +The study will record the exact model family, version, system prompts, retrieval configuration, source connectors, and confidence-scoring rules used during the intervention. Any model or prompt change during data collection will be logged as a protocol deviation and sensitivity-analysis flag. + +### 7.2 Prompt Taxonomy + +The mediator may generate five prompt types: + +| Prompt Type | Trigger | Required Recipient | +| --- | --- | --- | +| Action-item confirmation | Candidate who/what/when commitment extracted from communication | Named assignee | +| Decision confirmation | Scope, priority, acceptance criterion, or architecture decision detected | Product owner plus affected implementer | +| Blocker clarification | Blocker stated without owner, dependency, or next step | Blocked assignee and blocker owner when identifiable | +| Conflict resolution | Communication claim conflicts with Jira/Git status | Assignee plus relevant role based on artifact type | +| Documentation completion | Confirmed item lacks required metadata | Assignee or scrum master/team lead | + +Prompt content must include the extracted claim, source link, linked artifact, confidence level, suggested action, and available responses: accept, edit, reject, defer, or mark sensitive. + +### 7.3 Confidence and Escalation Rules + +The mediator will use predefined confidence bands. High-confidence, low-impact items may be batched. Medium-confidence items require explicit confirmation. Low-confidence items are logged for technical evaluation but do not trigger participant prompts unless sampled for manual review. High-impact items always require confirmation regardless of confidence. + +| Confidence Band | Operational Rule | +| --- | --- | +| High | Confidence >= 0.80 and no conflict detected; batch unless high-impact. | +| Medium | 0.50 <= confidence < 0.80; request confirmation before writeback. | +| Low | Confidence < 0.50; do not prompt by default; include in manual evaluation sample. | +| High-impact override | Scope, priority, ownership, due date, acceptance criteria, security/privacy, or release decision; require role-aware confirmation. | + +### 7.4 Technical Performance Evaluation + +A stratified random sample of communication events and AI-generated candidates will be manually coded by two independent coders. The gold sample will include accepted prompts, edited prompts, rejected prompts, ignored prompts, and low-confidence non-prompted candidates. Technical outcomes will include precision, recall where denominators can be estimated, F1 score, false-positive categories, false-negative categories, and disagreement resolution notes for extraction, artifact linking, and conflict detection. + +### 7.5 Human Oversight + +The AI does not independently decide project scope, task status, ownership, or acceptance criteria. For low-risk summaries, one assignee confirmation may be sufficient. For high-impact updates, confirmation may be required from multiple roles, such as developer, QA, and product owner. + +### 7.6 Writeback Policy + +Confirmed updates may be written to Jira comments, issue metadata, pull request descriptions, decision records, or a team knowledge store. Writeback must preserve provenance. Every AI-created record should include source links, timestamp, confirming roles, and reversal instructions. + +The mediator may write comments or draft suggestions automatically after confirmation, but it may not silently change issue status, assignee, due date, sprint scope, acceptance criteria, or release labels. Those fields require explicit role-aware confirmation and a reversible audit trail. + +### 7.7 Prompt Governance and Participant Controls + +Prompt frequency will be capped to reduce interruption burden. The default cap is two prompts per participant per workday, excluding urgent high-impact conflicts. Participants may pause prompts for a defined period, mark a source as sensitive, reject a prompt without justification, edit the proposed record, request deletion from the study dataset when allowed by policy, or appeal a writeback to the data steward. The system should batch low-risk suggestions and prioritize prompts involving conflict, missing ownership, missing due date, unresolved blocker, or high-impact decision. + +## 8. Comparator + +The comparator is each team's baseline workflow without AI-mediated extraction, confirmation, or writeback. Teams continue to use their existing communication channels, issue trackers, stand-ups, and version-control practices. + +## 9. Outcomes + +### 9.1 Primary Outcome + +The primary outcome is change in Team Transparency Index (TTI) from baseline to intervention. + +```text +TTI = 0.25*COV + 0.25*CON + 0.20*CSN + 0.15*TML + 0.15*CMP +``` + +These weights are theory-informed a priori weights and will be fixed before data collection. Any change to the weights, component definitions, or component inclusion rules will require a documented protocol amendment and will not be tuned on outcome data. + +| Component | Operational Definition | +| --- | --- | +| COV | Proportion of eligible communication-mentioned tasks or decisions linked to a Jira issue, Git artifact, or decision record within the sprint window. Denominator excludes social talk, duplicate mentions, and explicitly out-of-scope personal content. | +| CON | Proportion of sampled status, ownership, blocker, and decision claims that match structured project records or are explicitly reconciled. | +| CSN | Proportion of high-impact decisions that meet the predefined role-confirmation threshold before writeback. | +| TML | Normalized inverse delay between event occurrence and documented update, capped at the sprint boundary. | +| CMP | Proportion of eligible action items with who, what, and when fields. "When" may be a due date, sprint, next meeting, or explicit "no date yet" confirmation. | + +TTI will be coded at sprint level for each team. The sampling window is the sprint plus a 48-hour post-sprint reconciliation period for records that are updated immediately after review or retrospective discussion. Eligible communication events are stand-up statements, chat messages, issue comments, pull request comments, and meeting transcript segments that contain a task, blocker, status claim, ownership claim, decision, due-date claim, or acceptance-criteria claim. Events are excluded when they are duplicate reminders, social conversation, private personnel content, or communication from non-consenting participants that cannot be de-identified under the approved protocol. + +The denominator for each component is constructed independently. For example, a statement such as "Ana will add rate-limit handling before release candidate 2" contributes to COV if it can be linked to an issue or pull request, to CON if the claim matches project records or is explicitly reconciled, to TML based on the time until the linked record is updated, and to CMP if the responsible person, work item, and timing are present. A scope decision such as "we are deferring SSO to the next sprint" contributes to CSN if it meets the predefined product-owner and affected-implementer confirmation rule. + +Two coders will independently code a 20% stratified sample of events during Stage A, covering each data source, role, prompt type, and confidence band. Disagreements will be adjudicated by a third reviewer. The study will report component-level reliability and will freeze the coding manual before Stage B only if the reliability progression criterion is met. + +### 9.2 Secondary Outcomes + +| Outcome | Measurement | +| --- | --- | +| Mean time to detect inconsistency | Time from first conflicting signal to system or human identification. | +| Documentation completeness | Share of action items and decisions with complete metadata. | +| Prompt burden | Number of AI prompts per participant per workday and participant-rated interruption cost. | +| Trust in AI updates | Survey items assessing perceived accuracy, explainability, and control. | +| Workload | NASA-TLX or raw NASA-TLX adapted for subjective workload measurement (Hart, 2006). | +| Psychological safety | Team psychological safety survey based on Edmondson's construct (Edmondson, 1999). | +| Adoption behavior | Acceptance, edit, rejection, and ignore rates for AI suggestions. | +| Technical extraction accuracy | Precision, recall where estimable, and F1 for action/decision/blocker extraction. | +| Technical linking accuracy | Accuracy and false-link rate for Jira/Git/decision-record links. | +| Conflict-detection accuracy | Precision and false-negative categories against manually coded conflict samples. | +| Governance concerns | Interview-coded concerns about privacy, surveillance, accountability, and misuse. | + +## 10. Variables + +| Role | Variable | Measurement | Scale | +| --- | --- | --- | --- | +| Intervention | AI mediator active | Baseline = 0, intervention = 1 | Nominal | +| Primary DV | TTI | Weighted composite of five normalized components | Interval | +| Secondary DV | Inconsistency detection time | Hours from conflict creation to detection | Ratio | +| Secondary DV | Documentation completeness | Complete items / total action items | Ratio | +| Secondary DV | Prompt burden | Prompts per user per day; survey burden rating | Ratio / ordinal | +| Secondary DV | Workload | NASA-TLX or raw NASA-TLX score | Interval | +| Secondary DV | Psychological safety | Mean survey score | Interval | +| Control | Team size | Number of active team members | Ratio | +| Control | Sprint phase | Planning, execution, release, retrospective | Nominal | +| Control | Workload intensity | Issue count, pull request count, release deadline indicator | Ratio / nominal | +| Confound | Management pressure | Interview and survey indicators | Qualitative / ordinal | +| Confound | Tool maturity | Baseline completeness and issue hygiene | Interval | + +## 11. Instruments and Data Sources + +| Instrument or Source | Purpose | Notes | +| --- | --- | --- | +| Jira or issue tracker export | Issue status, assignee, transitions, comments, timestamps | Metadata and project records only unless consent allows content review. | +| Git hosting metadata | Commits, pull requests, reviews, branch references | Used for traceability linking and event timing. | +| Chat and stand-up transcripts | Candidate decisions, blockers, status claims, action items | Sensitive content redaction required. | +| AI prompt log | Prompt type, recipient role, source evidence, confidence band, response, confirmation outcome, override reason | Used for adoption, burden, and safety-gate analysis. | +| TTI extraction rubric | Standardized coding of COV, CON, CSN, TML, CMP | Frozen after Stage A if reliability criteria are met. | +| Technical gold sample | Manually coded sample of candidate action items, decisions, blockers, links, and conflicts | Used to estimate extraction, linking, and conflict-detection performance. | +| Perceived transparency survey | Participant perception of alignment and visibility | Administer baseline and intervention. | +| Trust in AI update survey | Explainability, source confidence, reversibility, perceived accuracy | Administer after intervention. | +| NASA-TLX or raw NASA-TLX | Subjective workload | Use consistently across phases. | +| Psychological safety survey | Team climate for interpersonal risk taking | Use validated or adapted items with permission where required. | +| Semi-structured interviews | Qualitative adoption and governance data | Conduct after intervention. | + +## 12. Data Collection Procedure + +### 12.1 Preparation + +The research team will obtain organizational approval, ethics approval where required, and informed consent. Tool integrations will be configured with least-privilege access. A data mapping exercise will identify which fields are needed for outcome measurement and which fields must be excluded or redacted. + +### 12.2 Baseline Phase + +During baseline, the AI mediator will not prompt participants or write back to project systems. Data will be collected passively from agreed sources to compute baseline TTI and secondary measures. Participants will complete baseline surveys on perceived transparency, workload, and psychological safety. + +### 12.3 Intervention Phase + +During intervention, the AI mediator will generate candidate knowledge items and role-aware prompts. Participants may accept, edit, reject, or ignore prompts. Confirmed items may be written back according to the writeback policy. System logs will record prompt type, source evidence, confirmation route, and outcome. + +### 12.4 Follow-up + +Participants will complete post-intervention surveys. A purposive subset of participants across roles will be invited for interviews. Interviews will focus on usefulness, trust, interruptions, missed cases, false positives, correction behavior, and privacy concerns. + +## 13. Analysis Plan + +### 13.1 Stage A Feasibility Analysis + +Stage A will be analyzed as a feasibility pilot. The primary outputs will be recruitment and retention rates, consent coverage by data source, missingness by variable, prompt burden, safety-gate events, TTI coding reliability, and technical performance against the manually coded gold sample. Progression to Stage B requires meeting the criteria in Section 6. If criteria are not met, the study will be reported as a feasibility study and the intervention or protocol will be revised before any confirmatory evaluation. + +### 13.2 Quantitative Effectiveness Analysis + +If Stage B proceeds, the primary analysis will compare TTI between baseline and intervention. If multiple teams are included, mixed-effects models should be used with phase as a fixed effect and team as a random effect. If the sample is too small or assumptions are not met, the analysis will report descriptive statistics, paired comparisons, effect sizes, and confidence intervals. + +Secondary outcomes will be analyzed as follows: + +| Outcome | Primary Test | Fallback | +| --- | --- | --- | +| TTI | Linear mixed-effects model | Wilcoxon signed-rank or descriptive effect size | +| Detection time | Survival or time-to-event model | Mann-Whitney U or paired non-parametric comparison | +| Documentation completeness | Logistic or beta regression | Proportion difference with confidence interval | +| Prompt burden | Poisson or negative binomial model | Descriptive rate comparison | +| Survey scales | Paired t-test or mixed model | Wilcoxon signed-rank | +| Adoption behavior | Acceptance/edit/rejection rates | Descriptive and role-stratified analysis | +| Technical performance | Precision, recall, F1, false-positive and false-negative review | Descriptive error taxonomy | + +Multiple comparisons will be treated as exploratory unless the study is powered confirmatorily. Survey scale reliability will be assessed with internal consistency when sample size permits. + +### 13.3 Missing Data and Partial Consent + +Missing data will be reported by variable, phase, team, role, and source system. Partial consent will be handled by excluding non-consenting participants' identifiable content from qualitative analysis and by using only aggregate or de-identified metadata where permitted by the approved consent protocol. If consent gaps prevent reliable team-level TTI computation, that team-period will be excluded from primary effectiveness analysis and retained only for feasibility reporting. Sensitivity analyses will compare complete-case results with analyses using available de-identified metadata where ethically and statistically appropriate. + +### 13.4 Qualitative Analysis + +Interview transcripts and observation notes will be analyzed using thematic analysis. The initial coding frame will include trust, usefulness, interruption cost, correction behavior, false positives, missed updates, role conflict, surveillance concern, and psychological safety. Two coders should independently code a subset of material and reconcile disagreements before coding the full dataset. + +### 13.5 Mixed-Method Integration + +Quantitative and qualitative findings will be integrated through joint displays. For example, teams with increased TTI but high prompt burden will be examined qualitatively to determine whether the transparency gain was acceptable. Teams with low AI adoption will be examined for trust, workflow fit, and governance barriers. + +## 14. Data Management + +Data will be minimized to the fields required for the study. Raw chat or transcript content should be redacted or summarized when possible. Identifiers will be pseudonymized before analysis. A linkage file, if required, will be stored separately with restricted access. Data storage will use encrypted institutional or organizational storage. Retention period should be defined before data collection and communicated in the consent form. + +The TTI should be reported at team level. Individual-level prompt response data may be needed for analysis, but it must not be used for performance evaluation. Any publication should aggregate or anonymize examples to prevent re-identification. + +Data access will be separated by role. The principal investigator may access the full approved research dataset; the data steward may access linkage and redaction files; the technical integration owner may access system logs needed for debugging but not interview material; and team representatives may review only aggregated disclosure summaries. Managers will receive team-level summaries only after aggregation and disclosure review. + +## 15. Ethics and Governance + +This study involves workplace communication and therefore creates privacy and power-differential risks. The following safeguards are required: + +1. Participation must be voluntary and based on informed consent. +2. Team members must know which channels and artifacts are included. +3. Managers must not receive individual-level transparency or prompt-response scores. +4. AI-generated updates must be visible, source-linked, and reversible. +5. Sensitive personal content must be excluded or redacted. +6. Participants must be able to challenge or correct AI-generated records. +7. Interview participation must be separated from management evaluation. +8. Participants must be able to pause prompts, mark content sensitive, reject or edit suggested updates, request deletion of erroneous AI-generated records, and appeal contested records through a named governance route. +9. Manager-facing dashboards must exclude individual prompt-response behavior, individual transparency scores, and any ranking or comparison of named participants. + +Psychological safety is both an outcome and an ethical constraint. Edmondson (1999) defines psychological safety as a shared belief that the team is safe for interpersonal risk taking. A transparency tool that makes people afraid to surface blockers would fail even if it improves documentation metrics. + +### 15.1 Safety Gates and Stopping Rules + +The intervention will be paused for governance review if any of the following occur: + +1. Mean psychological safety drops by 0.5 or more on a five-point scale from baseline. +2. Mean raw NASA-TLX workload increases by more than 10 points from baseline. +3. Median prompt burden exceeds two prompts per participant per workday for two consecutive weeks. +4. More than 20% of participants use pause, opt-out, or mark-sensitive controls during a sprint. +5. Any participant reports perceived retaliation, coercion, or manager misuse linked to AI-mediated records. +6. Sensitive data are captured outside the approved source scope and are not remediated within the incident response window. + +TTI improvement will be considered acceptable only if these safety thresholds remain within bounds. A team that improves TTI while breaching safety thresholds will be reported as a governance failure rather than an effectiveness success. + +## 16. Risk Management + +| Risk | Likelihood | Impact | Mitigation | +| --- | --- | --- | --- | +| Excessive prompts interrupt work | Medium | Medium | Prompt caps, batching, priority rules, pause controls, safety-gate review. | +| False positives reduce trust | Medium | Medium | Source links, confidence labels, easy rejection/editing. | +| Surveillance perception | Medium | High | Consent, team-level reporting, no individual scoring, governance review. | +| Sensitive data captured | Medium | High | Channel scoping, redaction, access controls, data minimization. | +| Tool integration failure | Medium | Medium | Pilot mapping, fallback export, manual coding sample. | +| Management misuse | Low to medium | High | Written policy forbidding individual performance use. | +| AI writeback error | Medium | Medium | Human confirmation, reversible updates, audit trail, deletion request and appeal route. | + +## 17. Dissemination Plan + +Findings will be disseminated as a protocol paper, a field evaluation paper after data collection, and a practitioner-oriented report for participating teams. The field evaluation report will separate confirmed findings from exploratory observations and will disclose limitations related to sample size, team context, and tool configuration. + +Before recruitment, the study team will select a preregistration destination and artifact repository, such as OSF or an institutional repository, and will specify which protocol materials, de-identified analysis code, instrument templates, and non-sensitive aggregate outputs can be shared. The final field evaluation report will include a reporting checklist adapted from protocol-reporting and AI-intervention reporting guidance. + +## 18. Protocol Status + +This is a revised protocol draft prepared for preregistration and ethics review. It should not be treated as preregistered, ethics-approved, or implementation-ready until the following items are completed: + +1. Participating organization and teams. +2. Exact tool integrations and data fields. +3. Consent language. +4. Survey instruments and permissions. +5. Prompt governance thresholds. +6. Statistical analysis plan details based on expected team count and sprint duration. + +## 19. Appendix Roadmap + +The final preregistration package should include the following study artifacts: + +1. Consent form and participant information sheet. +2. Data source map and data dictionary. +3. TTI coding manual and adjudication guide. +4. Prompt taxonomy and prompt examples. +5. Survey items for perceived transparency, trust in AI updates, workload, and psychological safety. +6. Semi-structured interview guide. +7. Technical gold-sample coding guide. +8. Safety incident form and escalation workflow. +9. Statistical analysis plan and analysis code plan. + +## References + +Ball, K. (2021). *Electronic monitoring and surveillance in the workplace: Literature review and policy recommendations*. Publications Office of the European Union. https://doi.org/10.2760/451453 + +Chan, A.-W., Tetzlaff, J. M., Altman, D. G., Laupacis, A., Gotzsche, P. C., Krleza-Jeric, K., Hrobjartsson, A., Mann, H., Dickersin, K., Berlin, J. A., Dore, C. J., Parulekar, W. R., Summerskill, W. S. M., Groves, T., Schulz, K. F., Sox, H. C., Rockhold, F. W., Rennie, D., & Moher, D. (2013). SPIRIT 2013 statement: Defining standard protocol items for clinical trials. *Annals of Internal Medicine, 158*(3), 200-207. https://doi.org/10.7326/0003-4819-158-3-201302050-00583 + +Cinkusz, K., Chudziak, J. A., & Niewiadomska-Szynkiewicz, E. (2025). Cognitive agents powered by large language models for agile software project management. *Electronics, 14*(1), Article 87. https://doi.org/10.3390/electronics14010087 + +Cleland-Huang, J., Gotel, O. C. Z., Huffman Hayes, J., Mäder, P., & Zisman, A. (2014). Software traceability: Trends and future directions. *Future of Software Engineering Proceedings*, 55-69. https://doi.org/10.1145/2593882.2593891 + +Edmondson, A. (1999). Psychological safety and learning behavior in work teams. *Administrative Science Quarterly, 44*(2), 350-383. https://doi.org/10.2307/2666999 + +Hart, S. G. (2006). NASA-Task Load Index (NASA-TLX); 20 years later. *Proceedings of the Human Factors and Ergonomics Society Annual Meeting, 50*(9), 904-908. https://doi.org/10.1177/154193120605000909 + +Itzik, D., & Roy, G. (2023). Does agile methodology fit all characteristics of software projects? Review and analysis. *Empirical Software Engineering, 28*, Article 105. https://doi.org/10.1007/s10664-023-10334-7 + +Liu, X., Cruz Rivera, S., Moher, D., Calvert, M. J., Denniston, A. K., & SPIRIT-AI and CONSORT-AI Working Group. (2020). Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: The CONSORT-AI extension. *Nature Medicine, 26*, 1364-1374. https://doi.org/10.1038/s41591-020-1034-x + +Malla, P. (2025). Analyzing the impact of agile methodologies on software quality and delivery speed: A comparative study. *World Journal of Advanced Research and Reviews, 25*(1), 1207-1216. https://doi.org/10.30574/wjarr.2025.25.1.0184 + +Parasuraman, R., & Riley, V. (1997). Humans and automation: Use, misuse, disuse, abuse. *Human Factors, 39*(2), 230-253. https://doi.org/10.1518/001872097778543886 + +Shneiderman, B. (2020). Human-centered artificial intelligence: Reliable, safe & trustworthy. *International Journal of Human-Computer Interaction, 36*(6), 495-504. https://doi.org/10.1080/10447318.2020.1741118 + +Stray, V., Moe, N. B., & Bergersen, G. R. (2017). Are daily stand-up meetings valuable? A survey of developers in software teams. In H. Baumeister, H. Lichter, & M. Riebisch (Eds.), *Agile Processes in Software Engineering and Extreme Programming* (pp. 274-281). Springer. https://doi.org/10.1007/978-3-319-57633-6_20 + +Stray, V., Moe, N. B., & Sjoberg, D. I. K. (2020). Daily stand-up meetings: Start breaking the rules. *IEEE Software, 37*(3), 70-77. https://doi.org/10.1109/MS.2018.2875988 + +Stray, V., Sjoberg, D. I. K., & Dyba, T. (2016). The daily stand-up meeting: A grounded theory study. *Journal of Systems and Software, 114*, 101-124. https://doi.org/10.1016/j.jss.2016.01.004 + +Umar, M. A. M. A., Lano, K., & Abubakar, A. K. (2025). Automated requirements engineering framework for agile model-driven development. *Frontiers in Computer Science, 7*, Article 1537100. https://doi.org/10.3389/fcomp.2025.1537100 + +Verwijs, C., & Russo, D. (2024). Do Agile scaling approaches make a difference? An empirical comparison of team effectiveness across popular scaling approaches. *Empirical Software Engineering, 29*, Article 75. https://doi.org/10.1007/s10664-024-10481-5 diff --git a/Aidaily_final_manuscript.rtf b/Aidaily_final_manuscript.rtf new file mode 100644 index 0000000..a08594d --- /dev/null +++ b/Aidaily_final_manuscript.rtf @@ -0,0 +1,441 @@ +{\rtf1\ansi\ansicpg1252\cocoartf2870 +\cocoatextscaling0\cocoaplatform0{\fonttbl\f0\fswiss\fcharset0 Helvetica-Light;} +{\colortbl;\red255\green255\blue255;} +{\*\expandedcolortbl;;} +\pard\tx560\tx1120\tx1680\tx2240\tx2800\tx3360\tx3920\tx4480\tx5040\tx5600\tx6160\tx6720\pardirnatural\partightenfactor0 + +\f0\fs24 \cf0 # Protocol for Evaluating a Conversational AI Framework for Agile Team Transparency and Knowledge Traceability\ +\ +## Abstract\ +\ +**Background:** Agile teams rely on daily coordination, issue tracking, and version control to maintain shared understanding. In distributed and hybrid teams, however, decisions, blockers, and action commitments often remain fragmented across meetings, chat, Jira, and Git. Conversational AI may help by extracting candidate updates from informal communication, linking them to project artifacts, and prompting team members to confirm or correct them.\ +\ +**Objective:** This protocol describes a mixed-method field study to evaluate whether a conversational AI framework improves Agile team transparency, knowledge traceability, and perceived alignment without increasing cognitive burden or weakening psychological safety.\ +\ +**Methods:** The study will use a two-stage mixed-method design. Stage A is a feasibility pilot that estimates measurement reliability, prompt burden, extraction/linking accuracy, and safety signals. Stage B is an optional powered field evaluation using a quasi-experimental, baseline-to-intervention design across Agile software teams. During baseline, teams continue normal workflows while communication, issue, and version-control metadata are measured. During intervention, teams use a conversational AI mediator that ingests stand-up transcripts, chat messages, Jira data, and Git metadata; extracts candidate decisions and action items; detects inconsistencies; and requests role-aware confirmation before writeback. The primary feasibility outcome is whether the Team Transparency Index (TTI) can be computed reliably. The primary effectiveness outcome, if Stage B proceeds, is change in TTI from baseline to intervention.\ +\ +**Analysis:** Feasibility analysis will report reliability, prompt-burden rates, technical performance, missingness, and progression criteria. Effectiveness analysis will compare baseline and intervention periods using mixed-effects models or non-parametric alternatives when assumptions are not met. Qualitative interviews and observation notes will be analyzed thematically to explain adoption patterns, trust, interruption costs, and governance concerns.\ +\ +**Ethics and Dissemination:** The study requires informed consent, role-based access controls, provenance-preserving audit trails, and safeguards against individual performance scoring. Results will be reported as a protocol-compliant field evaluation and will distinguish system performance from team-level organizational outcomes.\ +\ +**Keywords:** Agile software development; protocol paper; conversational AI; mixed methods; field study; team transparency; knowledge traceability; psychological safety; human-AI collaboration\ +\ +## 1. Introduction\ +\ +Agile software development depends on shared context. Teams coordinate through daily stand-ups, chat threads, issue trackers, code reviews, and version-control activity. These artifacts are individually useful, but they do not automatically produce a durable and verified team memory. A decision may be made verbally, partly clarified in chat, reflected indirectly in a pull request, and never updated in the issue tracker. The result is a persistent gap between what the team knows informally and what the system of record says formally.\ +\ +Prior research on daily stand-ups shows both the value and limits of recurring Agile communication. Stand-ups are widely used and can support team awareness, but their value varies by team size, role, and meeting quality (Stray et al., 2017). Grounded theory work also shows that stand-ups support coordination and monitoring while remaining sensitive to local practice (Stray et al., 2016). Later work argues that teams should adapt stand-up rules when the ritual no longer serves communication needs (Stray et al., 2020). These findings suggest that Agile transparency cannot be inferred from ceremony adoption alone.\ +\ +Research on Agile scaling and methodology fit also supports a context-sensitive approach. Verwijs and Russo (2024) found that scaling frameworks themselves explain little practical difference in team effectiveness. Itzik and Roy (2023) argue that Agile fit depends on software project characteristics and should be assessed through a decision framework. For a transparency intervention, this means the target is not framework compliance but the quality of alignment among people, artifacts, and decisions.\ +\ +AI-supported software project management provides relevant technical foundations. Automated requirements engineering research shows that machine learning can extract structured models from natural-language requirements in Agile contexts (Umar et al., 2025). LLM-based multi-agent project-management frameworks such as CogniSim show that AI agents can support Agile roles and project workflows in simulated environments (Cinkusz et al., 2025). Comparative work on Agile methods and technology-enhanced practices also suggests that AI-enabled tools may affect delivery and quality outcomes, while introducing risks of over-reliance on automation (Malla, 2025). The traceability literature is also directly relevant: software traceability is valuable but often performed ad hoc and after the fact, limiting its realized benefit (Cleland-Huang et al., 2014). Human-centered AI research further emphasizes that reliable and trustworthy systems should combine high automation with meaningful human control (Shneiderman, 2020). These strands motivate a field study that tests whether conversational AI can improve the human communication layer of real Agile teams without producing automation bias, intrusive monitoring, or reduced psychological safety (Parasuraman & Riley, 1997; Ball, 2021).\ +\ +This protocol defines a study to evaluate a conversational AI mediator for Agile transparency. The protocol is informed by general protocol-reporting principles for transparent intervention studies, including the logic of SPIRIT-style completeness and AI-specific reporting attention to intervention behavior, human oversight, and error handling (Chan et al., 2013; Liu et al., 2020). The present study is not a clinical trial, so these guidelines are used as structural inspiration rather than as formal regulatory requirements.\ +\ +## 2. Study Objectives\ +\ +### 2.1 Primary Objective\ +\ +To determine whether use of a conversational AI mediator improves team-level transparency, as measured by change in Team Transparency Index (TTI) from baseline to intervention.\ +\ +### 2.2 Secondary Objectives\ +\ +1. To estimate whether TTI can be coded with acceptable inter-rater reliability in real Agile work artifacts.\ +2. To estimate the technical performance of AI extraction, artifact linking, and conflict detection against a manually coded gold sample.\ +3. To determine whether the intervention reduces mean time to detect communication-to-record inconsistencies.\ +4. To determine whether the intervention improves documentation completeness for action items and decisions.\ +5. To assess whether the intervention changes perceived transparency, cognitive workload, trust in AI-generated updates, and psychological safety.\ +6. To characterize qualitative adoption patterns, including when teams accept, correct, ignore, or reject AI-generated prompts.\ +7. To identify governance risks associated with AI-mediated team memory, especially surveillance concerns and misuse of transparency metrics for individual evaluation.\ +\ +## 3. Research Questions and Hypotheses\ +\ +**RQ1:** Does the conversational AI mediator improve team transparency compared with baseline practice?\ +\ +**H1a:** In the feasibility pilot, TTI components will reach acceptable coding reliability, defined as Cohen's kappa or Krippendorff's alpha >= 0.70 for categorical judgments and intraclass correlation >= 0.70 for continuous timing measures.\ +\ +**H1b:** In the powered field evaluation, mean TTI will be higher during the intervention period than during the baseline period.\ +\ +**RQ2:** Does the mediator improve detection of mismatches between team communication and project records?\ +\ +**H2:** Mean time to detect communication-to-record inconsistencies will be lower during intervention than during baseline.\ +\ +**RQ3:** Does the mediator improve documentation quality without increasing perceived burden?\ +\ +**H3a:** Documentation completeness will increase during intervention.\ +\ +**H3b:** Perceived workload will not increase by more than 10 points on a 0-100 raw NASA-TLX scale, and AI prompts will not exceed a median of two prompts per participant per workday.\ +\ +**RQ4:** How do team members experience AI-mediated confirmation prompts?\ +\ +**H4:** Team members will report higher trust in AI-generated updates when prompts include source links, confidence levels, and reversible writeback.\ +\ +**RQ5:** What governance safeguards are necessary to preserve psychological safety?\ +\ +This question is exploratory and will be answered through interviews, observations, and thematic analysis.\ +\ +## 4. Study Design\ +\ +The study will use a two-stage mixed-method design.\ +\ +**Stage A: Feasibility pilot.** Stage A tests whether the protocol can be implemented safely and whether TTI can be measured reliably. It estimates coding reliability, AI technical performance, prompt burden, missingness, opt-out rates, and safety signals. Stage A is not powered to test effectiveness.\ +\ +**Stage B: Field evaluation.** Stage B proceeds only if Stage A meets progression criteria. It uses a quasi-experimental repeated-measures design in which each participating team completes a baseline phase followed by an intervention phase. If organizational scheduling permits, teams will be staggered so that not all teams begin the intervention simultaneously. Staggering improves interpretability by separating intervention effects from calendar events such as release deadlines or organizational changes.\ +\ +The recommended minimum duration is:\ +\ +| Phase | Duration | Purpose |\ +| --- | --- | --- |\ +| Preparation | 2 weeks | Consent, tool configuration, privacy review, pilot data mapping |\ +| Stage A baseline | 1 sprint or 2 weeks | Test passive data capture and initial TTI coding |\ +| Stage A intervention | 1 sprint or 2 weeks | Test prompts, technical accuracy, safety gates, and burden |\ +| Stage A decision | 1 week | Apply progression criteria before Stage B |\ +| Stage B baseline | 2 sprints or 4 weeks | Measure normal workflow without AI writeback |\ +| Stage B intervention | 2 to 4 sprints or 4 to 8 weeks | Deploy AI mediator with confirmation prompts |\ +| Follow-up | 1 to 2 weeks | Interviews, debrief, data-quality checks |\ +\ +The design is not blinded. Participants will know when the AI mediator is active. Outcome extraction from system logs should be automated where possible and reviewed using predefined rules to reduce subjective bias.\ +\ +## 5. Setting\ +\ +The study will be conducted in software development teams that use Agile practices and maintain digital project artifacts. The minimum tooling environment is:\ +\ +1. An issue tracker such as Jira.\ +2. A Git-based version-control system.\ +3. A team communication channel such as Rocket.Chat, Slack, Microsoft Teams, or equivalent.\ +4. Recurring stand-up communication, either synchronous or asynchronous.\ +\ +The initial pilot may use a single Scrum team of 6 to 10 members. A stronger field evaluation should include at least 6 teams to support team-level comparison and reduce the risk that findings reflect one team's habits. Baseline tooling must be documented for each team, including existing bots, meeting summarizers, Jira automation rules, issue-linking practices, and dashboard use.\ +\ +## 6. Participants\ +\ +### 6.1 Target Population\ +\ +Participants are members of Agile software development teams, including developers, QA engineers, product owners, scrum masters, engineering managers, and other roles who participate in daily coordination or issue updates.\ +\ +### 6.2 Inclusion Criteria\ +\ +1. The participant is a member of a participating Agile team.\ +2. The participant uses the team's issue tracker, code review system, or communication channel as part of normal work.\ +3. The participant is at least 18 years old.\ +4. The participant provides informed consent for study data collection.\ +\ +### 6.3 Exclusion Criteria\ +\ +1. Participants who do not consent to data collection.\ +2. Contractors or external stakeholders whose communication cannot be captured under the organization's data policy.\ +3. Team members whose role creates a direct power conflict that cannot be mitigated in consent or interview procedures.\ +\ +### 6.4 Sampling Strategy\ +\ +Team recruitment will use purposive sampling. The study should prioritize teams with active project work, regular stand-up practices, and enough tool usage to support traceability measurement. Within recruited teams, all eligible members should be invited to participate to reduce selection bias.\ +\ +### 6.5 Target Sample Size\ +\ +For Stage A, the target is 1 to 3 teams and approximately 8 to 30 participants. Stage A will be judged by feasibility rather than statistical significance. Progression to Stage B requires: (a) TTI coding reliability >= 0.70, (b) median prompt burden <= 2 prompts per participant per workday, (c) no unresolved ethics or safety gate breach, (d) AI extraction/linking precision >= 0.70 on the manually coded pilot sample, and (e) no more than 20% missingness in the primary data streams.\ +\ +For Stage B, the target should be at least 6 teams if feasible. A final power analysis will be completed after Stage A using the observed TTI variance, estimated intraclass correlation, team count, sprint count, and expected missingness. If the available team count is too small for confirmatory inference, Stage B will be reported as an expanded feasibility and estimation study rather than an effectiveness trial.\ +\ +## 7. Intervention\ +\ +### 7.1 Conversational AI Mediator\ +\ +The intervention is a conversational AI framework embedded in the team's communication and project-management environment. It performs four functions:\ +\ +1. **Ingestion:** Collects meeting transcripts, chat messages, Jira updates, and Git metadata.\ +2. **Extraction and linking:** Identifies candidate action items, decisions, blockers, status claims, and links to project artifacts.\ +3. **Conflict detection:** Flags mismatches between communication and project records.\ +4. **Role-aware confirmation:** Prompts relevant team members to confirm, reject, or edit candidate knowledge artifacts before writeback.\ +\ +The study will record the exact model family, version, system prompts, retrieval configuration, source connectors, and confidence-scoring rules used during the intervention. Any model or prompt change during data collection will be logged as a protocol deviation and sensitivity-analysis flag.\ +\ +### 7.2 Prompt Taxonomy\ +\ +The mediator may generate five prompt types:\ +\ +| Prompt Type | Trigger | Required Recipient |\ +| --- | --- | --- |\ +| Action-item confirmation | Candidate who/what/when commitment extracted from communication | Named assignee |\ +| Decision confirmation | Scope, priority, acceptance criterion, or architecture decision detected | Product owner plus affected implementer |\ +| Blocker clarification | Blocker stated without owner, dependency, or next step | Blocked assignee and blocker owner when identifiable |\ +| Conflict resolution | Communication claim conflicts with Jira/Git status | Assignee plus relevant role based on artifact type |\ +| Documentation completion | Confirmed item lacks required metadata | Assignee or scrum master/team lead |\ +\ +Prompt content must include the extracted claim, source link, linked artifact, confidence level, suggested action, and available responses: accept, edit, reject, defer, or mark sensitive.\ +\ +### 7.3 Confidence and Escalation Rules\ +\ +The mediator will use predefined confidence bands. High-confidence, low-impact items may be batched. Medium-confidence items require explicit confirmation. Low-confidence items are logged for technical evaluation but do not trigger participant prompts unless sampled for manual review. High-impact items always require confirmation regardless of confidence.\ +\ +| Confidence Band | Operational Rule |\ +| --- | --- |\ +| High | Confidence >= 0.80 and no conflict detected; batch unless high-impact. |\ +| Medium | 0.50 <= confidence < 0.80; request confirmation before writeback. |\ +| Low | Confidence < 0.50; do not prompt by default; include in manual evaluation sample. |\ +| High-impact override | Scope, priority, ownership, due date, acceptance criteria, security/privacy, or release decision; require role-aware confirmation. |\ +\ +### 7.4 Technical Performance Evaluation\ +\ +A stratified random sample of communication events and AI-generated candidates will be manually coded by two independent coders. The gold sample will include accepted prompts, edited prompts, rejected prompts, ignored prompts, and low-confidence non-prompted candidates. Technical outcomes will include precision, recall where denominators can be estimated, F1 score, false-positive categories, false-negative categories, and disagreement resolution notes for extraction, artifact linking, and conflict detection.\ +\ +### 7.5 Human Oversight\ +\ +The AI does not independently decide project scope, task status, ownership, or acceptance criteria. For low-risk summaries, one assignee confirmation may be sufficient. For high-impact updates, confirmation may be required from multiple roles, such as developer, QA, and product owner.\ +\ +### 7.6 Writeback Policy\ +\ +Confirmed updates may be written to Jira comments, issue metadata, pull request descriptions, decision records, or a team knowledge store. Writeback must preserve provenance. Every AI-created record should include source links, timestamp, confirming roles, and reversal instructions.\ +\ +The mediator may write comments or draft suggestions automatically after confirmation, but it may not silently change issue status, assignee, due date, sprint scope, acceptance criteria, or release labels. Those fields require explicit role-aware confirmation and a reversible audit trail.\ +\ +### 7.7 Prompt Governance and Participant Controls\ +\ +Prompt frequency will be capped to reduce interruption burden. The default cap is two prompts per participant per workday, excluding urgent high-impact conflicts. Participants may pause prompts for a defined period, mark a source as sensitive, reject a prompt without justification, edit the proposed record, request deletion from the study dataset when allowed by policy, or appeal a writeback to the data steward. The system should batch low-risk suggestions and prioritize prompts involving conflict, missing ownership, missing due date, unresolved blocker, or high-impact decision.\ +\ +## 8. Comparator\ +\ +The comparator is each team's baseline workflow without AI-mediated extraction, confirmation, or writeback. Teams continue to use their existing communication channels, issue trackers, stand-ups, and version-control practices.\ +\ +## 9. Outcomes\ +\ +### 9.1 Primary Outcome\ +\ +The primary outcome is change in Team Transparency Index (TTI) from baseline to intervention.\ +\ +```text\ +TTI = 0.25*COV + 0.25*CON + 0.20*CSN + 0.15*TML + 0.15*CMP\ +```\ +\ +These weights are theory-informed a priori weights and will be fixed before data collection. Any change to the weights, component definitions, or component inclusion rules will require a documented protocol amendment and will not be tuned on outcome data.\ +\ +| Component | Operational Definition |\ +| --- | --- |\ +| COV | Proportion of eligible communication-mentioned tasks or decisions linked to a Jira issue, Git artifact, or decision record within the sprint window. Denominator excludes social talk, duplicate mentions, and explicitly out-of-scope personal content. |\ +| CON | Proportion of sampled status, ownership, blocker, and decision claims that match structured project records or are explicitly reconciled. |\ +| CSN | Proportion of high-impact decisions that meet the predefined role-confirmation threshold before writeback. |\ +| TML | Normalized inverse delay between event occurrence and documented update, capped at the sprint boundary. |\ +| CMP | Proportion of eligible action items with who, what, and when fields. "When" may be a due date, sprint, next meeting, or explicit "no date yet" confirmation. |\ +\ +TTI will be coded at sprint level for each team. The sampling window is the sprint plus a 48-hour post-sprint reconciliation period for records that are updated immediately after review or retrospective discussion. Eligible communication events are stand-up statements, chat messages, issue comments, pull request comments, and meeting transcript segments that contain a task, blocker, status claim, ownership claim, decision, due-date claim, or acceptance-criteria claim. Events are excluded when they are duplicate reminders, social conversation, private personnel content, or communication from non-consenting participants that cannot be de-identified under the approved protocol.\ +\ +The denominator for each component is constructed independently. For example, a statement such as "Ana will add rate-limit handling before release candidate 2" contributes to COV if it can be linked to an issue or pull request, to CON if the claim matches project records or is explicitly reconciled, to TML based on the time until the linked record is updated, and to CMP if the responsible person, work item, and timing are present. A scope decision such as "we are deferring SSO to the next sprint" contributes to CSN if it meets the predefined product-owner and affected-implementer confirmation rule.\ +\ +Two coders will independently code a 20% stratified sample of events during Stage A, covering each data source, role, prompt type, and confidence band. Disagreements will be adjudicated by a third reviewer. The study will report component-level reliability and will freeze the coding manual before Stage B only if the reliability progression criterion is met.\ +\ +### 9.2 Secondary Outcomes\ +\ +| Outcome | Measurement |\ +| --- | --- |\ +| Mean time to detect inconsistency | Time from first conflicting signal to system or human identification. |\ +| Documentation completeness | Share of action items and decisions with complete metadata. |\ +| Prompt burden | Number of AI prompts per participant per workday and participant-rated interruption cost. |\ +| Trust in AI updates | Survey items assessing perceived accuracy, explainability, and control. |\ +| Workload | NASA-TLX or raw NASA-TLX adapted for subjective workload measurement (Hart, 2006). |\ +| Psychological safety | Team psychological safety survey based on Edmondson's construct (Edmondson, 1999). |\ +| Adoption behavior | Acceptance, edit, rejection, and ignore rates for AI suggestions. |\ +| Technical extraction accuracy | Precision, recall where estimable, and F1 for action/decision/blocker extraction. |\ +| Technical linking accuracy | Accuracy and false-link rate for Jira/Git/decision-record links. |\ +| Conflict-detection accuracy | Precision and false-negative categories against manually coded conflict samples. |\ +| Governance concerns | Interview-coded concerns about privacy, surveillance, accountability, and misuse. |\ +\ +## 10. Variables\ +\ +| Role | Variable | Measurement | Scale |\ +| --- | --- | --- | --- |\ +| Intervention | AI mediator active | Baseline = 0, intervention = 1 | Nominal |\ +| Primary DV | TTI | Weighted composite of five normalized components | Interval |\ +| Secondary DV | Inconsistency detection time | Hours from conflict creation to detection | Ratio |\ +| Secondary DV | Documentation completeness | Complete items / total action items | Ratio |\ +| Secondary DV | Prompt burden | Prompts per user per day; survey burden rating | Ratio / ordinal |\ +| Secondary DV | Workload | NASA-TLX or raw NASA-TLX score | Interval |\ +| Secondary DV | Psychological safety | Mean survey score | Interval |\ +| Control | Team size | Number of active team members | Ratio |\ +| Control | Sprint phase | Planning, execution, release, retrospective | Nominal |\ +| Control | Workload intensity | Issue count, pull request count, release deadline indicator | Ratio / nominal |\ +| Confound | Management pressure | Interview and survey indicators | Qualitative / ordinal |\ +| Confound | Tool maturity | Baseline completeness and issue hygiene | Interval |\ +\ +## 11. Instruments and Data Sources\ +\ +| Instrument or Source | Purpose | Notes |\ +| --- | --- | --- |\ +| Jira or issue tracker export | Issue status, assignee, transitions, comments, timestamps | Metadata and project records only unless consent allows content review. |\ +| Git hosting metadata | Commits, pull requests, reviews, branch references | Used for traceability linking and event timing. |\ +| Chat and stand-up transcripts | Candidate decisions, blockers, status claims, action items | Sensitive content redaction required. |\ +| AI prompt log | Prompt type, recipient role, source evidence, confidence band, response, confirmation outcome, override reason | Used for adoption, burden, and safety-gate analysis. |\ +| TTI extraction rubric | Standardized coding of COV, CON, CSN, TML, CMP | Frozen after Stage A if reliability criteria are met. |\ +| Technical gold sample | Manually coded sample of candidate action items, decisions, blockers, links, and conflicts | Used to estimate extraction, linking, and conflict-detection performance. |\ +| Perceived transparency survey | Participant perception of alignment and visibility | Administer baseline and intervention. |\ +| Trust in AI update survey | Explainability, source confidence, reversibility, perceived accuracy | Administer after intervention. |\ +| NASA-TLX or raw NASA-TLX | Subjective workload | Use consistently across phases. |\ +| Psychological safety survey | Team climate for interpersonal risk taking | Use validated or adapted items with permission where required. |\ +| Semi-structured interviews | Qualitative adoption and governance data | Conduct after intervention. |\ +\ +## 12. Data Collection Procedure\ +\ +### 12.1 Preparation\ +\ +The research team will obtain organizational approval, ethics approval where required, and informed consent. Tool integrations will be configured with least-privilege access. A data mapping exercise will identify which fields are needed for outcome measurement and which fields must be excluded or redacted.\ +\ +### 12.2 Baseline Phase\ +\ +During baseline, the AI mediator will not prompt participants or write back to project systems. Data will be collected passively from agreed sources to compute baseline TTI and secondary measures. Participants will complete baseline surveys on perceived transparency, workload, and psychological safety.\ +\ +### 12.3 Intervention Phase\ +\ +During intervention, the AI mediator will generate candidate knowledge items and role-aware prompts. Participants may accept, edit, reject, or ignore prompts. Confirmed items may be written back according to the writeback policy. System logs will record prompt type, source evidence, confirmation route, and outcome.\ +\ +### 12.4 Follow-up\ +\ +Participants will complete post-intervention surveys. A purposive subset of participants across roles will be invited for interviews. Interviews will focus on usefulness, trust, interruptions, missed cases, false positives, correction behavior, and privacy concerns.\ +\ +## 13. Analysis Plan\ +\ +### 13.1 Stage A Feasibility Analysis\ +\ +Stage A will be analyzed as a feasibility pilot. The primary outputs will be recruitment and retention rates, consent coverage by data source, missingness by variable, prompt burden, safety-gate events, TTI coding reliability, and technical performance against the manually coded gold sample. Progression to Stage B requires meeting the criteria in Section 6. If criteria are not met, the study will be reported as a feasibility study and the intervention or protocol will be revised before any confirmatory evaluation.\ +\ +### 13.2 Quantitative Effectiveness Analysis\ +\ +If Stage B proceeds, the primary analysis will compare TTI between baseline and intervention. If multiple teams are included, mixed-effects models should be used with phase as a fixed effect and team as a random effect. If the sample is too small or assumptions are not met, the analysis will report descriptive statistics, paired comparisons, effect sizes, and confidence intervals.\ +\ +Secondary outcomes will be analyzed as follows:\ +\ +| Outcome | Primary Test | Fallback |\ +| --- | --- | --- |\ +| TTI | Linear mixed-effects model | Wilcoxon signed-rank or descriptive effect size |\ +| Detection time | Survival or time-to-event model | Mann-Whitney U or paired non-parametric comparison |\ +| Documentation completeness | Logistic or beta regression | Proportion difference with confidence interval |\ +| Prompt burden | Poisson or negative binomial model | Descriptive rate comparison |\ +| Survey scales | Paired t-test or mixed model | Wilcoxon signed-rank |\ +| Adoption behavior | Acceptance/edit/rejection rates | Descriptive and role-stratified analysis |\ +| Technical performance | Precision, recall, F1, false-positive and false-negative review | Descriptive error taxonomy |\ +\ +Multiple comparisons will be treated as exploratory unless the study is powered confirmatorily. Survey scale reliability will be assessed with internal consistency when sample size permits.\ +\ +### 13.3 Missing Data and Partial Consent\ +\ +Missing data will be reported by variable, phase, team, role, and source system. Partial consent will be handled by excluding non-consenting participants' identifiable content from qualitative analysis and by using only aggregate or de-identified metadata where permitted by the approved consent protocol. If consent gaps prevent reliable team-level TTI computation, that team-period will be excluded from primary effectiveness analysis and retained only for feasibility reporting. Sensitivity analyses will compare complete-case results with analyses using available de-identified metadata where ethically and statistically appropriate.\ +\ +### 13.4 Qualitative Analysis\ +\ +Interview transcripts and observation notes will be analyzed using thematic analysis. The initial coding frame will include trust, usefulness, interruption cost, correction behavior, false positives, missed updates, role conflict, surveillance concern, and psychological safety. Two coders should independently code a subset of material and reconcile disagreements before coding the full dataset.\ +\ +### 13.5 Mixed-Method Integration\ +\ +Quantitative and qualitative findings will be integrated through joint displays. For example, teams with increased TTI but high prompt burden will be examined qualitatively to determine whether the transparency gain was acceptable. Teams with low AI adoption will be examined for trust, workflow fit, and governance barriers.\ +\ +## 14. Data Management\ +\ +Data will be minimized to the fields required for the study. Raw chat or transcript content should be redacted or summarized when possible. Identifiers will be pseudonymized before analysis. A linkage file, if required, will be stored separately with restricted access. Data storage will use encrypted institutional or organizational storage. Retention period should be defined before data collection and communicated in the consent form.\ +\ +The TTI should be reported at team level. Individual-level prompt response data may be needed for analysis, but it must not be used for performance evaluation. Any publication should aggregate or anonymize examples to prevent re-identification.\ +\ +Data access will be separated by role. The principal investigator may access the full approved research dataset; the data steward may access linkage and redaction files; the technical integration owner may access system logs needed for debugging but not interview material; and team representatives may review only aggregated disclosure summaries. Managers will receive team-level summaries only after aggregation and disclosure review.\ +\ +## 15. Ethics and Governance\ +\ +This study involves workplace communication and therefore creates privacy and power-differential risks. The following safeguards are required:\ +\ +1. Participation must be voluntary and based on informed consent.\ +2. Team members must know which channels and artifacts are included.\ +3. Managers must not receive individual-level transparency or prompt-response scores.\ +4. AI-generated updates must be visible, source-linked, and reversible.\ +5. Sensitive personal content must be excluded or redacted.\ +6. Participants must be able to challenge or correct AI-generated records.\ +7. Interview participation must be separated from management evaluation.\ +8. Participants must be able to pause prompts, mark content sensitive, reject or edit suggested updates, request deletion of erroneous AI-generated records, and appeal contested records through a named governance route.\ +9. Manager-facing dashboards must exclude individual prompt-response behavior, individual transparency scores, and any ranking or comparison of named participants.\ +\ +Psychological safety is both an outcome and an ethical constraint. Edmondson (1999) defines psychological safety as a shared belief that the team is safe for interpersonal risk taking. A transparency tool that makes people afraid to surface blockers would fail even if it improves documentation metrics.\ +\ +### 15.1 Safety Gates and Stopping Rules\ +\ +The intervention will be paused for governance review if any of the following occur:\ +\ +1. Mean psychological safety drops by 0.5 or more on a five-point scale from baseline.\ +2. Mean raw NASA-TLX workload increases by more than 10 points from baseline.\ +3. Median prompt burden exceeds two prompts per participant per workday for two consecutive weeks.\ +4. More than 20% of participants use pause, opt-out, or mark-sensitive controls during a sprint.\ +5. Any participant reports perceived retaliation, coercion, or manager misuse linked to AI-mediated records.\ +6. Sensitive data are captured outside the approved source scope and are not remediated within the incident response window.\ +\ +TTI improvement will be considered acceptable only if these safety thresholds remain within bounds. A team that improves TTI while breaching safety thresholds will be reported as a governance failure rather than an effectiveness success.\ +\ +## 16. Risk Management\ +\ +| Risk | Likelihood | Impact | Mitigation |\ +| --- | --- | --- | --- |\ +| Excessive prompts interrupt work | Medium | Medium | Prompt caps, batching, priority rules, pause controls, safety-gate review. |\ +| False positives reduce trust | Medium | Medium | Source links, confidence labels, easy rejection/editing. |\ +| Surveillance perception | Medium | High | Consent, team-level reporting, no individual scoring, governance review. |\ +| Sensitive data captured | Medium | High | Channel scoping, redaction, access controls, data minimization. |\ +| Tool integration failure | Medium | Medium | Pilot mapping, fallback export, manual coding sample. |\ +| Management misuse | Low to medium | High | Written policy forbidding individual performance use. |\ +| AI writeback error | Medium | Medium | Human confirmation, reversible updates, audit trail, deletion request and appeal route. |\ +\ +## 17. Dissemination Plan\ +\ +Findings will be disseminated as a protocol paper, a field evaluation paper after data collection, and a practitioner-oriented report for participating teams. The field evaluation report will separate confirmed findings from exploratory observations and will disclose limitations related to sample size, team context, and tool configuration.\ +\ +Before recruitment, the study team will select a preregistration destination and artifact repository, such as OSF or an institutional repository, and will specify which protocol materials, de-identified analysis code, instrument templates, and non-sensitive aggregate outputs can be shared. The final field evaluation report will include a reporting checklist adapted from protocol-reporting and AI-intervention reporting guidance.\ +\ +## 18. Protocol Status\ +\ +This is a revised protocol draft prepared for preregistration and ethics review. It should not be treated as preregistered, ethics-approved, or implementation-ready until the following items are completed:\ +\ +1. Participating organization and teams.\ +2. Exact tool integrations and data fields.\ +3. Consent language.\ +4. Survey instruments and permissions.\ +5. Prompt governance thresholds.\ +6. Statistical analysis plan details based on expected team count and sprint duration.\ +\ +## 19. Appendix Roadmap\ +\ +The final preregistration package should include the following study artifacts:\ +\ +1. Consent form and participant information sheet.\ +2. Data source map and data dictionary.\ +3. TTI coding manual and adjudication guide.\ +4. Prompt taxonomy and prompt examples.\ +5. Survey items for perceived transparency, trust in AI updates, workload, and psychological safety.\ +6. Semi-structured interview guide.\ +7. Technical gold-sample coding guide.\ +8. Safety incident form and escalation workflow.\ +9. Statistical analysis plan and analysis code plan.\ +\ +## References\ +\ +Ball, K. (2021). *Electronic monitoring and surveillance in the workplace: Literature review and policy recommendations*. Publications Office of the European Union. https://doi.org/10.2760/451453\ +\ +Chan, A.-W., Tetzlaff, J. M., Altman, D. G., Laupacis, A., Gotzsche, P. C., Krleza-Jeric, K., Hrobjartsson, A., Mann, H., Dickersin, K., Berlin, J. A., Dore, C. J., Parulekar, W. R., Summerskill, W. S. M., Groves, T., Schulz, K. F., Sox, H. C., Rockhold, F. W., Rennie, D., & Moher, D. (2013). SPIRIT 2013 statement: Defining standard protocol items for clinical trials. *Annals of Internal Medicine, 158*(3), 200-207. https://doi.org/10.7326/0003-4819-158-3-201302050-00583\ +\ +Cinkusz, K., Chudziak, J. A., & Niewiadomska-Szynkiewicz, E. (2025). Cognitive agents powered by large language models for agile software project management. *Electronics, 14*(1), Article 87. https://doi.org/10.3390/electronics14010087\ +\ +Cleland-Huang, J., Gotel, O. C. Z., Huffman Hayes, J., M\'e4der, P., & Zisman, A. (2014). Software traceability: Trends and future directions. *Future of Software Engineering Proceedings*, 55-69. https://doi.org/10.1145/2593882.2593891\ +\ +Edmondson, A. (1999). Psychological safety and learning behavior in work teams. *Administrative Science Quarterly, 44*(2), 350-383. https://doi.org/10.2307/2666999\ +\ +Hart, S. G. (2006). NASA-Task Load Index (NASA-TLX); 20 years later. *Proceedings of the Human Factors and Ergonomics Society Annual Meeting, 50*(9), 904-908. https://doi.org/10.1177/154193120605000909\ +\ +Itzik, D., & Roy, G. (2023). Does agile methodology fit all characteristics of software projects? Review and analysis. *Empirical Software Engineering, 28*, Article 105. https://doi.org/10.1007/s10664-023-10334-7\ +\ +Liu, X., Cruz Rivera, S., Moher, D., Calvert, M. J., Denniston, A. K., & SPIRIT-AI and CONSORT-AI Working Group. (2020). Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: The CONSORT-AI extension. *Nature Medicine, 26*, 1364-1374. https://doi.org/10.1038/s41591-020-1034-x\ +\ +Malla, P. (2025). Analyzing the impact of agile methodologies on software quality and delivery speed: A comparative study. *World Journal of Advanced Research and Reviews, 25*(1), 1207-1216. https://doi.org/10.30574/wjarr.2025.25.1.0184\ +\ +Parasuraman, R., & Riley, V. (1997). Humans and automation: Use, misuse, disuse, abuse. *Human Factors, 39*(2), 230-253. https://doi.org/10.1518/001872097778543886\ +\ +Shneiderman, B. (2020). Human-centered artificial intelligence: Reliable, safe & trustworthy. *International Journal of Human-Computer Interaction, 36*(6), 495-504. https://doi.org/10.1080/10447318.2020.1741118\ +\ +Stray, V., Moe, N. B., & Bergersen, G. R. (2017). Are daily stand-up meetings valuable? A survey of developers in software teams. In H. Baumeister, H. Lichter, & M. Riebisch (Eds.), *Agile Processes in Software Engineering and Extreme Programming* (pp. 274-281). Springer. https://doi.org/10.1007/978-3-319-57633-6_20\ +\ +Stray, V., Moe, N. B., & Sjoberg, D. I. K. (2020). Daily stand-up meetings: Start breaking the rules. *IEEE Software, 37*(3), 70-77. https://doi.org/10.1109/MS.2018.2875988\ +\ +Stray, V., Sjoberg, D. I. K., & Dyba, T. (2016). The daily stand-up meeting: A grounded theory study. *Journal of Systems and Software, 114*, 101-124. https://doi.org/10.1016/j.jss.2016.01.004\ +\ +Umar, M. A. M. A., Lano, K., & Abubakar, A. K. (2025). Automated requirements engineering framework for agile model-driven development. *Frontiers in Computer Science, 7*, Article 1537100. https://doi.org/10.3389/fcomp.2025.1537100\ +\ +Verwijs, C., & Russo, D. (2024). Do Agile scaling approaches make a difference? An empirical comparison of team effectiveness across popular scaling approaches. *Empirical Software Engineering, 29*, Article 75. https://doi.org/10.1007/s10664-024-10481-5\ +} \ No newline at end of file diff --git a/Aidaily_final_manuscript_standalone_review.md b/Aidaily_final_manuscript_standalone_review.md new file mode 100644 index 0000000..43d7a31 --- /dev/null +++ b/Aidaily_final_manuscript_standalone_review.md @@ -0,0 +1,160 @@ +# Standalone Academic Paper Review: Aidaily Final Manuscript + +## Manuscript Reviewed + +- File: `Aidaily_final_manuscript.md` +- Title: Protocol for Evaluating a Conversational AI Framework for Agile Team Transparency and Knowledge Traceability +- Word count: 5,774 +- Review mode: academic-paper-reviewer / full standalone review + +## Reviewer Configuration + +| Role | Configured Identity | Review Focus | +| --- | --- | --- | +| Editor-in-Chief | Editor for an empirical software engineering / HCI methods venue | Contribution, venue fit, protocol completeness, publishability | +| Reviewer 1 | Mixed-methods field-study methodologist | Design, measurement, sampling, analysis, feasibility-to-effectiveness transition | +| Reviewer 2 | Agile software engineering and traceability scholar | Literature fit, software engineering contribution, traceability framing | +| Reviewer 3 | Human-centered AI and workplace governance reviewer | Human-AI interaction, consent, surveillance risk, participant protections | +| Devil's Advocate | Adversarial protocol reviewer | Strongest rejection arguments, hidden assumptions, failure modes | + +## Editorial Decision + +Minor Revision. + +The manuscript is suitable as a protocol paper after targeted finalization work. It no longer overclaims empirical findings, has a defensible two-stage study design, and gives unusually explicit attention to AI governance and participant controls. The remaining issues are submission-readiness issues rather than fundamental design flaws. + +## EIC Review + +### Strengths + +1. The manuscript is clearly framed as a prospective protocol rather than a completed empirical evaluation. +2. The two-stage design is appropriate: Stage A establishes feasibility, reliability, safety, and technical performance before Stage B attempts effectiveness inference. +3. The primary contribution is coherent: a protocol for evaluating conversational AI as a mediator of Agile team memory and traceability. +4. The paper has a credible ethical stance, especially the rule that improved TTI is not an effectiveness success if psychological safety or workload thresholds fail. + +### Required Minor Revisions + +1. The paper should state its intended venue category more explicitly in the introduction or final paragraph: empirical software engineering protocol, HCI field-study protocol, or AI governance intervention protocol. +2. The appendix roadmap is useful, but a submission-ready protocol should include at least abbreviated versions of key appendices: TTI coding rubric, prompt examples, interview guide, and survey item sources. +3. The Protocol Status section is honest, but it may read as too unfinished for some journals. Reframe it as "Items to complete before recruitment" rather than a limitation of manuscript readiness. + +### EIC Recommendation + +Accept after minor revision for a protocol-friendly venue; otherwise revise format and appendices for the target journal. + +## Methodology Review + +### Strengths + +1. The feasibility-to-effectiveness separation is sound. +2. The TTI is now operationalized with eligibility rules, component denominators, examples, coder sampling, and reliability thresholds. +3. Missing-data and partial-consent handling are addressed at a protocol level. +4. The analysis plan correctly avoids promising confirmatory inference if Stage B is underpowered. + +### Weaknesses + +1. The TTI remains a newly proposed composite index. The manuscript fixes the weights a priori, but it should still justify why those weights reflect transparency rather than documentation hygiene. +2. The Stage B power-analysis plan is deferred until Stage A. That is acceptable, but the manuscript should include a minimum analyzable unit, such as minimum team-periods or sprint observations required to report Stage B as effectiveness rather than expanded feasibility. +3. The qualitative analysis plan names thematic analysis but does not specify whether the coding approach is reflexive, codebook, framework, or hybrid. + +### Required Minor Revisions + +1. Add a short rationale for TTI component weights and clarify that sensitivity analyses will report unweighted or component-level outcomes. +2. Add minimum criteria for calling Stage B an effectiveness evaluation. +3. Specify the qualitative coding approach and how interview findings will be integrated with safety-gate interpretation. + +## Domain Review + +### Strengths + +1. The paper is well positioned against Agile ceremonies, traceability, AI project-management support, and human-centered AI. +2. The baseline characterization requirement is important and domain-appropriate. +3. The distinction between improving artifact traceability and improving team transparency is visible throughout the manuscript. + +### Weaknesses + +1. The literature base is adequate but still compact. A final target-journal submission may need more work on coordination theory, team cognition, software bots, meeting summarization, and traceability recovery. +2. The comparator is "normal workflow," but the manuscript could explain more clearly how baseline automation will be classified and compared across teams. +3. The proposed intervention overlaps with existing tooling such as issue bots, meeting summarizers, AI assistants, and traceability recommenders. The novelty claim should more sharply identify what is new: role-aware confirmation plus governance-gated writeback plus team-level transparency measurement. + +### Required Minor Revisions + +1. Add a concise novelty paragraph distinguishing the framework from ordinary summarization, project-management bots, and issue-linking automation. +2. Add a baseline automation taxonomy or table for existing team tools. +3. Expand the literature by 3-6 sources if submitting to an empirical software engineering venue. + +## Human-Centered AI and Governance Review + +### Strengths + +1. Participant controls are concrete: pause, sensitive marking, reject, edit, delete request, and appeal. +2. The manager dashboard restrictions are strong and necessary. +3. The safety gates are actionable and include both quantitative and qualitative harm signals. +4. The protocol avoids treating AI acceptance as correctness. + +### Weaknesses + +1. The system interface is still abstract. The paper describes prompt content but not prompt timing, batching UI, timeout behavior, or how disagreements between roles appear to users. +2. "Mark sensitive" and deletion requests need more operational detail: who receives the request, what is removed from the study dataset, what remains in organizational systems, and what audit trail is retained. +3. The appeal route names a data steward but does not specify escalation timing or independence from management. + +### Required Minor Revisions + +1. Add 2-3 example prompt templates or interaction states. +2. Add a short data-rights workflow for sensitive marking, deletion request, appeal, and retained audit metadata. +3. Specify maximum response time for governance review after a safety-gate trigger. + +## Devil's Advocate Review + +### Strongest Rejection Argument + +The protocol may still conflate "better documented work" with "more transparent teamwork." TTI is substantially artifact-oriented: links, consistency, confirmation, timeliness, and metadata completeness. A team could improve all five dimensions while becoming less candid in informal communication because participants know their statements may become project records. The manuscript mitigates this with psychological safety, workload, opt-out, and governance gates, but the core construct-validity concern remains and should be directly acknowledged. + +### Stress-Test Issues + +1. If team members move sensitive coordination into private channels, the system may improve visible traceability while degrading actual transparency. +2. If managers pressure teams to accept prompts, confirmation logs may become compliance artifacts rather than consent signals. +3. If the AI mediator performs poorly for ambiguous social or product decisions, the technical accuracy metrics may look acceptable on action items while missing the hardest transparency cases. +4. If Stage A is run in a highly cooperative team, progression criteria may not generalize to teams with lower psychological safety. + +### Required Minor Revisions + +1. Add construct-validity limitations for TTI. +2. Add a qualitative probe about communication displacement into private channels. +3. Stratify technical performance by prompt type, not only overall extraction/linking/conflict detection. + +## Consolidated Revision Roadmap + +### Must Fix Before Submission + +| # | Issue | Location | Required Action | +| --- | --- | --- | --- | +| 1 | TTI construct validity needs more explicit limitation and rationale. | Sections 9, 13, limitations/status area | Explain why weights were chosen, report component-level sensitivity analyses, and acknowledge artifact-transparency limitations. | +| 2 | Study artifacts are roadmap-only. | Appendix Roadmap | Add abbreviated appendices or supplementary-file placeholders for TTI rubric, prompt examples, surveys, and interview guide. | +| 3 | Novelty relative to existing bots/summarizers is implicit. | Introduction / Intervention | Add a novelty paragraph distinguishing role-aware confirmation and governance-gated writeback. | +| 4 | Governance workflows need operational detail. | Ethics / Data Management | Specify sensitive marking, deletion, appeal, safety-gate review timing, and retained audit metadata. | + +### Should Fix + +| # | Issue | Suggested Action | +| --- | --- | --- | +| 5 | Qualitative analysis approach is generic. | Specify codebook, reflexive, framework, or hybrid thematic analysis. | +| 6 | Stage B effectiveness threshold is underdefined. | Add minimum team-period or sprint-observation criteria. | +| 7 | Interface details are abstract. | Add prompt templates and timeout/batching behavior. | +| 8 | Baseline automation varies by team. | Add a baseline automation classification table. | + +## Publication Readiness Score + +| Dimension | Score | Rationale | +| --- | --- | --- | +| Contribution clarity | 8/10 | Strong protocol contribution; novelty can be sharper. | +| Methodological rigor | 8/10 | Good feasibility/effectiveness separation; TTI construct validity needs more discussion. | +| Ethical/governance completeness | 8.5/10 | Strong safeguards; operational data-rights workflow should be added. | +| Literature grounding | 7/10 | Adequate for draft; should expand for target venue. | +| Submission polish | 7/10 | Clean manuscript, but appendices and target-venue formatting remain. | + +Overall readiness: 7.7/10. + +## Final Recommendation + +Minor Revision. The manuscript is credible and close to submission as a protocol paper. The most important final edits are to add artifact appendices, sharpen novelty, and make TTI's construct-validity limitations explicit. No major redesign is required. diff --git a/Aidaily_protocol_paper.md b/Aidaily_protocol_paper.md new file mode 100644 index 0000000..af1816f --- /dev/null +++ b/Aidaily_protocol_paper.md @@ -0,0 +1,459 @@ +# Protocol for Evaluating a Conversational AI Framework for Agile Team Transparency and Knowledge Traceability + +## Material Passport + +- Origin Skill: academic-research-suite / experiment-agent +- Origin Mode: protocol paper / human study protocol +- Origin Date: 2026-06-26 +- Verification Status: STAGE_4_5_FINAL_INTEGRITY_PASS +- Version Label: aidaily_protocol_paper_v3_integrity_pass +- Source Draft: Aidaily_v0.3_revised.md + +## Abstract + +**Background:** Agile teams rely on daily coordination, issue tracking, and version control to maintain shared understanding. In distributed and hybrid teams, however, decisions, blockers, and action commitments often remain fragmented across meetings, chat, Jira, and Git. Conversational AI may help by extracting candidate updates from informal communication, linking them to project artifacts, and prompting team members to confirm or correct them. + +**Objective:** This protocol describes a mixed-method field study to evaluate whether a conversational AI framework improves Agile team transparency, knowledge traceability, and perceived alignment without increasing cognitive burden or weakening psychological safety. + +**Methods:** The study will use a two-stage mixed-method design. Stage A is a feasibility pilot that estimates measurement reliability, prompt burden, extraction/linking accuracy, and safety signals. Stage B is an optional powered field evaluation using a quasi-experimental, baseline-to-intervention design across Agile software teams. During baseline, teams continue normal workflows while communication, issue, and version-control metadata are measured. During intervention, teams use a conversational AI mediator that ingests stand-up transcripts, chat messages, Jira data, and Git metadata; extracts candidate decisions and action items; detects inconsistencies; and requests role-aware confirmation before writeback. The primary feasibility outcome is whether the Team Transparency Index (TTI) can be computed reliably. The primary effectiveness outcome, if Stage B proceeds, is change in TTI from baseline to intervention. + +**Analysis:** Feasibility analysis will report reliability, prompt-burden rates, technical performance, missingness, and progression criteria. Effectiveness analysis will compare baseline and intervention periods using mixed-effects models or non-parametric alternatives when assumptions are not met. Qualitative interviews and observation notes will be analyzed thematically to explain adoption patterns, trust, interruption costs, and governance concerns. + +**Ethics and Dissemination:** The study requires informed consent, role-based access controls, provenance-preserving audit trails, and safeguards against individual performance scoring. Results will be reported as a protocol-compliant field evaluation and will distinguish system performance from team-level organizational outcomes. + +**Keywords:** Agile software development; protocol paper; conversational AI; mixed methods; field study; team transparency; knowledge traceability; psychological safety; human-AI collaboration + +## 1. Introduction + +Agile software development depends on shared context. Teams coordinate through daily stand-ups, chat threads, issue trackers, code reviews, and version-control activity. These artifacts are individually useful, but they do not automatically produce a durable and verified team memory. A decision may be made verbally, partly clarified in chat, reflected indirectly in a pull request, and never updated in the issue tracker. The result is a persistent gap between what the team knows informally and what the system of record says formally. + +Prior research on daily stand-ups shows both the value and limits of recurring Agile communication. Stand-ups are widely used and can support team awareness, but their value varies by team size, role, and meeting quality (Stray et al., 2017). Grounded theory work also shows that stand-ups support coordination and monitoring while remaining sensitive to local practice (Stray et al., 2016). Later work argues that teams should adapt stand-up rules when the ritual no longer serves communication needs (Stray et al., 2020). These findings suggest that Agile transparency cannot be inferred from ceremony adoption alone. + +Research on Agile scaling and methodology fit also supports a context-sensitive approach. Verwijs and Russo (2024) found that scaling frameworks themselves explain little practical difference in team effectiveness. Itzik and Roy (2023) argue that Agile fit depends on software project characteristics and should be assessed through a decision framework. For a transparency intervention, this means the target is not framework compliance but the quality of alignment among people, artifacts, and decisions. + +AI-supported software project management provides relevant technical foundations. Automated requirements engineering research shows that machine learning can extract structured models from natural-language requirements in Agile contexts (Umar et al., 2025). LLM-based multi-agent project-management frameworks such as CogniSim show that AI agents can support Agile roles and project workflows in simulated environments (Cinkusz et al., 2025). Comparative work on Agile methods and technology-enhanced practices also suggests that AI-enabled tools may affect delivery and quality outcomes, while introducing risks of over-reliance on automation (Malla, 2025). The traceability literature is also directly relevant: software traceability is valuable but often performed ad hoc and after the fact, limiting its realized benefit (Cleland-Huang et al., 2014). Human-centered AI research further emphasizes that reliable and trustworthy systems should combine high automation with meaningful human control (Shneiderman, 2020). These strands motivate a field study that tests whether conversational AI can improve the human communication layer of real Agile teams without producing automation bias, intrusive monitoring, or reduced psychological safety (Parasuraman & Riley, 1997; Ball, 2021). + +This protocol defines a study to evaluate a conversational AI mediator for Agile transparency. The protocol is informed by general protocol-reporting principles for transparent intervention studies, including the logic of SPIRIT-style completeness and AI-specific reporting attention to intervention behavior, human oversight, and error handling (Chan et al., 2013; Liu et al., 2020). The present study is not a clinical trial, so these guidelines are used as structural inspiration rather than as formal regulatory requirements. + +## 2. Study Objectives + +### 2.1 Primary Objective + +To determine whether use of a conversational AI mediator improves team-level transparency, as measured by change in Team Transparency Index (TTI) from baseline to intervention. + +### 2.2 Secondary Objectives + +1. To estimate whether TTI can be coded with acceptable inter-rater reliability in real Agile work artifacts. +2. To estimate the technical performance of AI extraction, artifact linking, and conflict detection against a manually coded gold sample. +3. To determine whether the intervention reduces mean time to detect communication-to-record inconsistencies. +4. To determine whether the intervention improves documentation completeness for action items and decisions. +5. To assess whether the intervention changes perceived transparency, cognitive workload, trust in AI-generated updates, and psychological safety. +6. To characterize qualitative adoption patterns, including when teams accept, correct, ignore, or reject AI-generated prompts. +7. To identify governance risks associated with AI-mediated team memory, especially surveillance concerns and misuse of transparency metrics for individual evaluation. + +## 3. Research Questions and Hypotheses + +**RQ1:** Does the conversational AI mediator improve team transparency compared with baseline practice? + +**H1a:** In the feasibility pilot, TTI components will reach acceptable coding reliability, defined as Cohen's kappa or Krippendorff's alpha >= 0.70 for categorical judgments and intraclass correlation >= 0.70 for continuous timing measures. + +**H1b:** In the powered field evaluation, mean TTI will be higher during the intervention period than during the baseline period. + +**RQ2:** Does the mediator improve detection of mismatches between team communication and project records? + +**H2:** Mean time to detect communication-to-record inconsistencies will be lower during intervention than during baseline. + +**RQ3:** Does the mediator improve documentation quality without increasing perceived burden? + +**H3a:** Documentation completeness will increase during intervention. + +**H3b:** Perceived workload will not increase by more than 10 points on a 0-100 raw NASA-TLX scale, and AI prompts will not exceed a median of two prompts per participant per workday. + +**RQ4:** How do team members experience AI-mediated confirmation prompts? + +**H4:** Team members will report higher trust in AI-generated updates when prompts include source links, confidence levels, and reversible writeback. + +**RQ5:** What governance safeguards are necessary to preserve psychological safety? + +This question is exploratory and will be answered through interviews, observations, and thematic analysis. + +## 4. Study Design + +The study will use a two-stage mixed-method design. + +**Stage A: Feasibility pilot.** Stage A tests whether the protocol can be implemented safely and whether TTI can be measured reliably. It estimates coding reliability, AI technical performance, prompt burden, missingness, opt-out rates, and safety signals. Stage A is not powered to test effectiveness. + +**Stage B: Field evaluation.** Stage B proceeds only if Stage A meets progression criteria. It uses a quasi-experimental repeated-measures design in which each participating team completes a baseline phase followed by an intervention phase. If organizational scheduling permits, teams will be staggered so that not all teams begin the intervention simultaneously. Staggering improves interpretability by separating intervention effects from calendar events such as release deadlines or organizational changes. + +The recommended minimum duration is: + +| Phase | Duration | Purpose | +| --- | --- | --- | +| Preparation | 2 weeks | Consent, tool configuration, privacy review, pilot data mapping | +| Stage A baseline | 1 sprint or 2 weeks | Test passive data capture and initial TTI coding | +| Stage A intervention | 1 sprint or 2 weeks | Test prompts, technical accuracy, safety gates, and burden | +| Stage A decision | 1 week | Apply progression criteria before Stage B | +| Stage B baseline | 2 sprints or 4 weeks | Measure normal workflow without AI writeback | +| Stage B intervention | 2 to 4 sprints or 4 to 8 weeks | Deploy AI mediator with confirmation prompts | +| Follow-up | 1 to 2 weeks | Interviews, debrief, data-quality checks | + +The design is not blinded. Participants will know when the AI mediator is active. Outcome extraction from system logs should be automated where possible and reviewed using predefined rules to reduce subjective bias. + +## 5. Setting + +The study will be conducted in software development teams that use Agile practices and maintain digital project artifacts. The minimum tooling environment is: + +1. An issue tracker such as Jira. +2. A Git-based version-control system. +3. A team communication channel such as Rocket.Chat, Slack, Microsoft Teams, or equivalent. +4. Recurring stand-up communication, either synchronous or asynchronous. + +The initial pilot may use a single Scrum team of 6 to 10 members. A stronger field evaluation should include at least 6 teams to support team-level comparison and reduce the risk that findings reflect one team's habits. Baseline tooling must be documented for each team, including existing bots, meeting summarizers, Jira automation rules, issue-linking practices, and dashboard use. + +## 6. Participants + +### 6.1 Target Population + +Participants are members of Agile software development teams, including developers, QA engineers, product owners, scrum masters, engineering managers, and other roles who participate in daily coordination or issue updates. + +### 6.2 Inclusion Criteria + +1. The participant is a member of a participating Agile team. +2. The participant uses the team's issue tracker, code review system, or communication channel as part of normal work. +3. The participant is at least 18 years old. +4. The participant provides informed consent for study data collection. + +### 6.3 Exclusion Criteria + +1. Participants who do not consent to data collection. +2. Contractors or external stakeholders whose communication cannot be captured under the organization's data policy. +3. Team members whose role creates a direct power conflict that cannot be mitigated in consent or interview procedures. + +### 6.4 Sampling Strategy + +Team recruitment will use purposive sampling. The study should prioritize teams with active project work, regular stand-up practices, and enough tool usage to support traceability measurement. Within recruited teams, all eligible members should be invited to participate to reduce selection bias. + +### 6.5 Target Sample Size + +For Stage A, the target is 1 to 3 teams and approximately 8 to 30 participants. Stage A will be judged by feasibility rather than statistical significance. Progression to Stage B requires: (a) TTI coding reliability >= 0.70, (b) median prompt burden <= 2 prompts per participant per workday, (c) no unresolved ethics or safety gate breach, (d) AI extraction/linking precision >= 0.70 on the manually coded pilot sample, and (e) no more than 20% missingness in the primary data streams. + +For Stage B, the target should be at least 6 teams if feasible. A final power analysis will be completed after Stage A using the observed TTI variance, estimated intraclass correlation, team count, sprint count, and expected missingness. If the available team count is too small for confirmatory inference, Stage B will be reported as an expanded feasibility and estimation study rather than an effectiveness trial. + +## 7. Intervention + +### 7.1 Conversational AI Mediator + +The intervention is a conversational AI framework embedded in the team's communication and project-management environment. It performs four functions: + +1. **Ingestion:** Collects meeting transcripts, chat messages, Jira updates, and Git metadata. +2. **Extraction and linking:** Identifies candidate action items, decisions, blockers, status claims, and links to project artifacts. +3. **Conflict detection:** Flags mismatches between communication and project records. +4. **Role-aware confirmation:** Prompts relevant team members to confirm, reject, or edit candidate knowledge artifacts before writeback. + +The study will record the exact model family, version, system prompts, retrieval configuration, source connectors, and confidence-scoring rules used during the intervention. Any model or prompt change during data collection will be logged as a protocol deviation and sensitivity-analysis flag. + +### 7.2 Prompt Taxonomy + +The mediator may generate five prompt types: + +| Prompt Type | Trigger | Required Recipient | +| --- | --- | --- | +| Action-item confirmation | Candidate who/what/when commitment extracted from communication | Named assignee | +| Decision confirmation | Scope, priority, acceptance criterion, or architecture decision detected | Product owner plus affected implementer | +| Blocker clarification | Blocker stated without owner, dependency, or next step | Blocked assignee and blocker owner when identifiable | +| Conflict resolution | Communication claim conflicts with Jira/Git status | Assignee plus relevant role based on artifact type | +| Documentation completion | Confirmed item lacks required metadata | Assignee or scrum master/team lead | + +Prompt content must include the extracted claim, source link, linked artifact, confidence level, suggested action, and available responses: accept, edit, reject, defer, or mark sensitive. + +### 7.3 Confidence and Escalation Rules + +The mediator will use predefined confidence bands. High-confidence, low-impact items may be batched. Medium-confidence items require explicit confirmation. Low-confidence items are logged for technical evaluation but do not trigger participant prompts unless sampled for manual review. High-impact items always require confirmation regardless of confidence. + +| Confidence Band | Operational Rule | +| --- | --- | +| High | Confidence >= 0.80 and no conflict detected; batch unless high-impact. | +| Medium | 0.50 <= confidence < 0.80; request confirmation before writeback. | +| Low | Confidence < 0.50; do not prompt by default; include in manual evaluation sample. | +| High-impact override | Scope, priority, ownership, due date, acceptance criteria, security/privacy, or release decision; require role-aware confirmation. | + +### 7.4 Technical Performance Evaluation + +A stratified random sample of communication events and AI-generated candidates will be manually coded by two independent coders. The gold sample will include accepted prompts, edited prompts, rejected prompts, ignored prompts, and low-confidence non-prompted candidates. Technical outcomes will include precision, recall where denominators can be estimated, F1 score, false-positive categories, false-negative categories, and disagreement resolution notes for extraction, artifact linking, and conflict detection. + +### 7.5 Human Oversight + +The AI does not independently decide project scope, task status, ownership, or acceptance criteria. For low-risk summaries, one assignee confirmation may be sufficient. For high-impact updates, confirmation may be required from multiple roles, such as developer, QA, and product owner. + +### 7.6 Writeback Policy + +Confirmed updates may be written to Jira comments, issue metadata, pull request descriptions, decision records, or a team knowledge store. Writeback must preserve provenance. Every AI-created record should include source links, timestamp, confirming roles, and reversal instructions. + +The mediator may write comments or draft suggestions automatically after confirmation, but it may not silently change issue status, assignee, due date, sprint scope, acceptance criteria, or release labels. Those fields require explicit role-aware confirmation and a reversible audit trail. + +### 7.7 Prompt Governance and Participant Controls + +Prompt frequency will be capped to reduce interruption burden. The default cap is two prompts per participant per workday, excluding urgent high-impact conflicts. Participants may pause prompts for a defined period, mark a source as sensitive, reject a prompt without justification, edit the proposed record, request deletion from the study dataset when allowed by policy, or appeal a writeback to the data steward. The system should batch low-risk suggestions and prioritize prompts involving conflict, missing ownership, missing due date, unresolved blocker, or high-impact decision. + +## 8. Comparator + +The comparator is each team's baseline workflow without AI-mediated extraction, confirmation, or writeback. Teams continue to use their existing communication channels, issue trackers, stand-ups, and version-control practices. + +## 9. Outcomes + +### 9.1 Primary Outcome + +The primary outcome is change in Team Transparency Index (TTI) from baseline to intervention. + +```text +TTI = 0.25*COV + 0.25*CON + 0.20*CSN + 0.15*TML + 0.15*CMP +``` + +These weights are theory-informed a priori weights and will be fixed before data collection. Any change to the weights, component definitions, or component inclusion rules will require a documented protocol amendment and will not be tuned on outcome data. + +| Component | Operational Definition | +| --- | --- | +| COV | Proportion of eligible communication-mentioned tasks or decisions linked to a Jira issue, Git artifact, or decision record within the sprint window. Denominator excludes social talk, duplicate mentions, and explicitly out-of-scope personal content. | +| CON | Proportion of sampled status, ownership, blocker, and decision claims that match structured project records or are explicitly reconciled. | +| CSN | Proportion of high-impact decisions that meet the predefined role-confirmation threshold before writeback. | +| TML | Normalized inverse delay between event occurrence and documented update, capped at the sprint boundary. | +| CMP | Proportion of eligible action items with who, what, and when fields. "When" may be a due date, sprint, next meeting, or explicit "no date yet" confirmation. | + +TTI will be coded at sprint level for each team. The sampling window is the sprint plus a 48-hour post-sprint reconciliation period for records that are updated immediately after review or retrospective discussion. Eligible communication events are stand-up statements, chat messages, issue comments, pull request comments, and meeting transcript segments that contain a task, blocker, status claim, ownership claim, decision, due-date claim, or acceptance-criteria claim. Events are excluded when they are duplicate reminders, social conversation, private personnel content, or communication from non-consenting participants that cannot be de-identified under the approved protocol. + +The denominator for each component is constructed independently. For example, a statement such as "Ana will add rate-limit handling before release candidate 2" contributes to COV if it can be linked to an issue or pull request, to CON if the claim matches project records or is explicitly reconciled, to TML based on the time until the linked record is updated, and to CMP if the responsible person, work item, and timing are present. A scope decision such as "we are deferring SSO to the next sprint" contributes to CSN if it meets the predefined product-owner and affected-implementer confirmation rule. + +Two coders will independently code a 20% stratified sample of events during Stage A, covering each data source, role, prompt type, and confidence band. Disagreements will be adjudicated by a third reviewer. The study will report component-level reliability and will freeze the coding manual before Stage B only if the reliability progression criterion is met. + +### 9.2 Secondary Outcomes + +| Outcome | Measurement | +| --- | --- | +| Mean time to detect inconsistency | Time from first conflicting signal to system or human identification. | +| Documentation completeness | Share of action items and decisions with complete metadata. | +| Prompt burden | Number of AI prompts per participant per workday and participant-rated interruption cost. | +| Trust in AI updates | Survey items assessing perceived accuracy, explainability, and control. | +| Workload | NASA-TLX or raw NASA-TLX adapted for subjective workload measurement (Hart, 2006). | +| Psychological safety | Team psychological safety survey based on Edmondson's construct (Edmondson, 1999). | +| Adoption behavior | Acceptance, edit, rejection, and ignore rates for AI suggestions. | +| Technical extraction accuracy | Precision, recall where estimable, and F1 for action/decision/blocker extraction. | +| Technical linking accuracy | Accuracy and false-link rate for Jira/Git/decision-record links. | +| Conflict-detection accuracy | Precision and false-negative categories against manually coded conflict samples. | +| Governance concerns | Interview-coded concerns about privacy, surveillance, accountability, and misuse. | + +## 10. Variables + +| Role | Variable | Measurement | Scale | +| --- | --- | --- | --- | +| Intervention | AI mediator active | Baseline = 0, intervention = 1 | Nominal | +| Primary DV | TTI | Weighted composite of five normalized components | Interval | +| Secondary DV | Inconsistency detection time | Hours from conflict creation to detection | Ratio | +| Secondary DV | Documentation completeness | Complete items / total action items | Ratio | +| Secondary DV | Prompt burden | Prompts per user per day; survey burden rating | Ratio / ordinal | +| Secondary DV | Workload | NASA-TLX or raw NASA-TLX score | Interval | +| Secondary DV | Psychological safety | Mean survey score | Interval | +| Control | Team size | Number of active team members | Ratio | +| Control | Sprint phase | Planning, execution, release, retrospective | Nominal | +| Control | Workload intensity | Issue count, pull request count, release deadline indicator | Ratio / nominal | +| Confound | Management pressure | Interview and survey indicators | Qualitative / ordinal | +| Confound | Tool maturity | Baseline completeness and issue hygiene | Interval | + +## 11. Instruments and Data Sources + +| Instrument or Source | Purpose | Notes | +| --- | --- | --- | +| Jira or issue tracker export | Issue status, assignee, transitions, comments, timestamps | Metadata and project records only unless consent allows content review. | +| Git hosting metadata | Commits, pull requests, reviews, branch references | Used for traceability linking and event timing. | +| Chat and stand-up transcripts | Candidate decisions, blockers, status claims, action items | Sensitive content redaction required. | +| AI prompt log | Prompt type, recipient role, source evidence, confidence band, response, confirmation outcome, override reason | Used for adoption, burden, and safety-gate analysis. | +| TTI extraction rubric | Standardized coding of COV, CON, CSN, TML, CMP | Frozen after Stage A if reliability criteria are met. | +| Technical gold sample | Manually coded sample of candidate action items, decisions, blockers, links, and conflicts | Used to estimate extraction, linking, and conflict-detection performance. | +| Perceived transparency survey | Participant perception of alignment and visibility | Administer baseline and intervention. | +| Trust in AI update survey | Explainability, source confidence, reversibility, perceived accuracy | Administer after intervention. | +| NASA-TLX or raw NASA-TLX | Subjective workload | Use consistently across phases. | +| Psychological safety survey | Team climate for interpersonal risk taking | Use validated or adapted items with permission where required. | +| Semi-structured interviews | Qualitative adoption and governance data | Conduct after intervention. | + +## 12. Data Collection Procedure + +### 12.1 Preparation + +The research team will obtain organizational approval, ethics approval where required, and informed consent. Tool integrations will be configured with least-privilege access. A data mapping exercise will identify which fields are needed for outcome measurement and which fields must be excluded or redacted. + +### 12.2 Baseline Phase + +During baseline, the AI mediator will not prompt participants or write back to project systems. Data will be collected passively from agreed sources to compute baseline TTI and secondary measures. Participants will complete baseline surveys on perceived transparency, workload, and psychological safety. + +### 12.3 Intervention Phase + +During intervention, the AI mediator will generate candidate knowledge items and role-aware prompts. Participants may accept, edit, reject, or ignore prompts. Confirmed items may be written back according to the writeback policy. System logs will record prompt type, source evidence, confirmation route, and outcome. + +### 12.4 Follow-up + +Participants will complete post-intervention surveys. A purposive subset of participants across roles will be invited for interviews. Interviews will focus on usefulness, trust, interruptions, missed cases, false positives, correction behavior, and privacy concerns. + +## 13. Analysis Plan + +### 13.1 Stage A Feasibility Analysis + +Stage A will be analyzed as a feasibility pilot. The primary outputs will be recruitment and retention rates, consent coverage by data source, missingness by variable, prompt burden, safety-gate events, TTI coding reliability, and technical performance against the manually coded gold sample. Progression to Stage B requires meeting the criteria in Section 6. If criteria are not met, the study will be reported as a feasibility study and the intervention or protocol will be revised before any confirmatory evaluation. + +### 13.2 Quantitative Effectiveness Analysis + +If Stage B proceeds, the primary analysis will compare TTI between baseline and intervention. If multiple teams are included, mixed-effects models should be used with phase as a fixed effect and team as a random effect. If the sample is too small or assumptions are not met, the analysis will report descriptive statistics, paired comparisons, effect sizes, and confidence intervals. + +Secondary outcomes will be analyzed as follows: + +| Outcome | Primary Test | Fallback | +| --- | --- | --- | +| TTI | Linear mixed-effects model | Wilcoxon signed-rank or descriptive effect size | +| Detection time | Survival or time-to-event model | Mann-Whitney U or paired non-parametric comparison | +| Documentation completeness | Logistic or beta regression | Proportion difference with confidence interval | +| Prompt burden | Poisson or negative binomial model | Descriptive rate comparison | +| Survey scales | Paired t-test or mixed model | Wilcoxon signed-rank | +| Adoption behavior | Acceptance/edit/rejection rates | Descriptive and role-stratified analysis | +| Technical performance | Precision, recall, F1, false-positive and false-negative review | Descriptive error taxonomy | + +Multiple comparisons will be treated as exploratory unless the study is powered confirmatorily. Survey scale reliability will be assessed with internal consistency when sample size permits. + +### 13.3 Missing Data and Partial Consent + +Missing data will be reported by variable, phase, team, role, and source system. Partial consent will be handled by excluding non-consenting participants' identifiable content from qualitative analysis and by using only aggregate or de-identified metadata where permitted by the approved consent protocol. If consent gaps prevent reliable team-level TTI computation, that team-period will be excluded from primary effectiveness analysis and retained only for feasibility reporting. Sensitivity analyses will compare complete-case results with analyses using available de-identified metadata where ethically and statistically appropriate. + +### 13.4 Qualitative Analysis + +Interview transcripts and observation notes will be analyzed using thematic analysis. The initial coding frame will include trust, usefulness, interruption cost, correction behavior, false positives, missed updates, role conflict, surveillance concern, and psychological safety. Two coders should independently code a subset of material and reconcile disagreements before coding the full dataset. + +### 13.5 Mixed-Method Integration + +Quantitative and qualitative findings will be integrated through joint displays. For example, teams with increased TTI but high prompt burden will be examined qualitatively to determine whether the transparency gain was acceptable. Teams with low AI adoption will be examined for trust, workflow fit, and governance barriers. + +## 14. Data Management + +Data will be minimized to the fields required for the study. Raw chat or transcript content should be redacted or summarized when possible. Identifiers will be pseudonymized before analysis. A linkage file, if required, will be stored separately with restricted access. Data storage will use encrypted institutional or organizational storage. Retention period should be defined before data collection and communicated in the consent form. + +The TTI should be reported at team level. Individual-level prompt response data may be needed for analysis, but it must not be used for performance evaluation. Any publication should aggregate or anonymize examples to prevent re-identification. + +Data access will be separated by role. The principal investigator may access the full approved research dataset; the data steward may access linkage and redaction files; the technical integration owner may access system logs needed for debugging but not interview material; and team representatives may review only aggregated disclosure summaries. Managers will receive team-level summaries only after aggregation and disclosure review. + +## 15. Ethics and Governance + +This study involves workplace communication and therefore creates privacy and power-differential risks. The following safeguards are required: + +1. Participation must be voluntary and based on informed consent. +2. Team members must know which channels and artifacts are included. +3. Managers must not receive individual-level transparency or prompt-response scores. +4. AI-generated updates must be visible, source-linked, and reversible. +5. Sensitive personal content must be excluded or redacted. +6. Participants must be able to challenge or correct AI-generated records. +7. Interview participation must be separated from management evaluation. +8. Participants must be able to pause prompts, mark content sensitive, reject or edit suggested updates, request deletion of erroneous AI-generated records, and appeal contested records through a named governance route. +9. Manager-facing dashboards must exclude individual prompt-response behavior, individual transparency scores, and any ranking or comparison of named participants. + +Psychological safety is both an outcome and an ethical constraint. Edmondson (1999) defines psychological safety as a shared belief that the team is safe for interpersonal risk taking. A transparency tool that makes people afraid to surface blockers would fail even if it improves documentation metrics. + +### 15.1 Safety Gates and Stopping Rules + +The intervention will be paused for governance review if any of the following occur: + +1. Mean psychological safety drops by 0.5 or more on a five-point scale from baseline. +2. Mean raw NASA-TLX workload increases by more than 10 points from baseline. +3. Median prompt burden exceeds two prompts per participant per workday for two consecutive weeks. +4. More than 20% of participants use pause, opt-out, or mark-sensitive controls during a sprint. +5. Any participant reports perceived retaliation, coercion, or manager misuse linked to AI-mediated records. +6. Sensitive data are captured outside the approved source scope and are not remediated within the incident response window. + +TTI improvement will be considered acceptable only if these safety thresholds remain within bounds. A team that improves TTI while breaching safety thresholds will be reported as a governance failure rather than an effectiveness success. + +## 16. Risk Management + +| Risk | Likelihood | Impact | Mitigation | +| --- | --- | --- | --- | +| Excessive prompts interrupt work | Medium | Medium | Prompt caps, batching, priority rules, pause controls, safety-gate review. | +| False positives reduce trust | Medium | Medium | Source links, confidence labels, easy rejection/editing. | +| Surveillance perception | Medium | High | Consent, team-level reporting, no individual scoring, governance review. | +| Sensitive data captured | Medium | High | Channel scoping, redaction, access controls, data minimization. | +| Tool integration failure | Medium | Medium | Pilot mapping, fallback export, manual coding sample. | +| Management misuse | Low to medium | High | Written policy forbidding individual performance use. | +| AI writeback error | Medium | Medium | Human confirmation, reversible updates, audit trail, deletion request and appeal route. | + +## 17. Dissemination Plan + +Findings will be disseminated as a protocol paper, a field evaluation paper after data collection, and a practitioner-oriented report for participating teams. The field evaluation report will separate confirmed findings from exploratory observations and will disclose limitations related to sample size, team context, and tool configuration. + +Before recruitment, the study team will select a preregistration destination and artifact repository, such as OSF or an institutional repository, and will specify which protocol materials, de-identified analysis code, instrument templates, and non-sensitive aggregate outputs can be shared. The final field evaluation report will include a reporting checklist adapted from protocol-reporting and AI-intervention reporting guidance. + +## 18. Protocol Status + +This is a revised protocol draft prepared for preregistration and ethics review. It should not be treated as preregistered, ethics-approved, or implementation-ready until the following items are completed: + +1. Participating organization and teams. +2. Exact tool integrations and data fields. +3. Consent language. +4. Survey instruments and permissions. +5. Prompt governance thresholds. +6. Statistical analysis plan details based on expected team count and sprint duration. + +## 19. Appendix Roadmap + +The final preregistration package should include the following study artifacts: + +1. Consent form and participant information sheet. +2. Data source map and data dictionary. +3. TTI coding manual and adjudication guide. +4. Prompt taxonomy and prompt examples. +5. Survey items for perceived transparency, trust in AI updates, workload, and psychological safety. +6. Semi-structured interview guide. +7. Technical gold-sample coding guide. +8. Safety incident form and escalation workflow. +9. Statistical analysis plan and analysis code plan. + +## 20. Stage 4 Revision Response Matrix + +| Concern ID | Reviewer Source | Action Taken | Manuscript Location | Status | +| --- | --- | --- | --- | --- | +| P1-1 | R1 / EIC | Added TTI eligibility rules, component denominators, examples, stratified double coding, adjudication, and reliability progression criteria. | Sections 3, 6.5, 9.1, 11, 13.1 | Resolved | +| P1-2 | R1 | Recast the study as a two-stage program with Stage A feasibility and optional Stage B effectiveness evaluation. | Abstract, Sections 3, 4, 6.5, 13.1-13.2 | Resolved | +| P1-3 | EIC / R3 | Added intervention reproducibility requirements, prompt taxonomy, confidence bands, escalation rules, writeback constraints, and participant response options. | Sections 7.1-7.7, 11, 12.3 | Resolved | +| P1-4 | Devil's Advocate / R1 | Added manually coded technical gold sample and technical performance outcomes for extraction, linking, and conflict detection. | Sections 2.2, 7.4, 9.2, 11, 13.1-13.2 | Resolved | +| P1-5 | R3 / Devil's Advocate | Added safety gates and stopping rules for psychological safety, workload, prompt burden, opt-out behavior, surveillance complaints, and sensitive-data incidents. | Sections 3, 13.1, 15.1, 16 | Resolved | +| P1-6 | R3 | Added pause, sensitive marking, reject, edit, delete request, and appeal controls; restricted manager dashboards. | Sections 7.7, 14, 15, 16 | Resolved | +| P2-7 | R1 | Added Stage A progression criteria and made powered inference conditional on observed variance, ICC, team count, sprint count, and missingness. | Sections 4, 6.5, 13.1-13.2 | Resolved | +| P2-8 | R1 | Defined the burden threshold as no more than a 10-point raw NASA-TLX increase and median prompt burden of at most two prompts per participant per workday. | Sections 3, 6.5, 7.7, 15.1 | Resolved | +| P2-9 | R2 | Required baseline characterization of bots, summarizers, Jira automation, issue-linking practices, and dashboards. | Sections 5, 8 | Resolved | +| P2-10 | R2 | Added traceability, human-centered AI, automation misuse, and workplace monitoring literature. | Sections 1 and References | Resolved | +| P2-11 | R1 | Added missing-data and partial-consent rules for incomplete logs, survey nonresponse, and consent gaps within teams. | Sections 13.3, 14 | Resolved | + +## References + +Ball, K. (2021). *Electronic monitoring and surveillance in the workplace: Literature review and policy recommendations*. Publications Office of the European Union. https://doi.org/10.2760/451453 + +Chan, A.-W., Tetzlaff, J. M., Altman, D. G., Laupacis, A., Gotzsche, P. C., Krleza-Jeric, K., Hrobjartsson, A., Mann, H., Dickersin, K., Berlin, J. A., Dore, C. J., Parulekar, W. R., Summerskill, W. S. M., Groves, T., Schulz, K. F., Sox, H. C., Rockhold, F. W., Rennie, D., & Moher, D. (2013). SPIRIT 2013 statement: Defining standard protocol items for clinical trials. *Annals of Internal Medicine, 158*(3), 200-207. https://doi.org/10.7326/0003-4819-158-3-201302050-00583 + +Cinkusz, K., Chudziak, J. A., & Niewiadomska-Szynkiewicz, E. (2025). Cognitive agents powered by large language models for agile software project management. *Electronics, 14*(1), Article 87. https://doi.org/10.3390/electronics14010087 + +Cleland-Huang, J., Gotel, O. C. Z., Huffman Hayes, J., Mäder, P., & Zisman, A. (2014). Software traceability: Trends and future directions. *Future of Software Engineering Proceedings*, 55-69. https://doi.org/10.1145/2593882.2593891 + +Edmondson, A. (1999). Psychological safety and learning behavior in work teams. *Administrative Science Quarterly, 44*(2), 350-383. https://doi.org/10.2307/2666999 + +Hart, S. G. (2006). NASA-Task Load Index (NASA-TLX); 20 years later. *Proceedings of the Human Factors and Ergonomics Society Annual Meeting, 50*(9), 904-908. https://doi.org/10.1177/154193120605000909 + +Itzik, D., & Roy, G. (2023). Does agile methodology fit all characteristics of software projects? Review and analysis. *Empirical Software Engineering, 28*, Article 105. https://doi.org/10.1007/s10664-023-10334-7 + +Liu, X., Cruz Rivera, S., Moher, D., Calvert, M. J., Denniston, A. K., & SPIRIT-AI and CONSORT-AI Working Group. (2020). Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: The CONSORT-AI extension. *Nature Medicine, 26*, 1364-1374. https://doi.org/10.1038/s41591-020-1034-x + +Malla, P. (2025). Analyzing the impact of agile methodologies on software quality and delivery speed: A comparative study. *World Journal of Advanced Research and Reviews, 25*(1), 1207-1216. https://doi.org/10.30574/wjarr.2025.25.1.0184 + +Parasuraman, R., & Riley, V. (1997). Humans and automation: Use, misuse, disuse, abuse. *Human Factors, 39*(2), 230-253. https://doi.org/10.1518/001872097778543886 + +Shneiderman, B. (2020). Human-centered artificial intelligence: Reliable, safe & trustworthy. *International Journal of Human-Computer Interaction, 36*(6), 495-504. https://doi.org/10.1080/10447318.2020.1741118 + +Stray, V., Moe, N. B., & Bergersen, G. R. (2017). Are daily stand-up meetings valuable? A survey of developers in software teams. In H. Baumeister, H. Lichter, & M. Riebisch (Eds.), *Agile Processes in Software Engineering and Extreme Programming* (pp. 274-281). Springer. https://doi.org/10.1007/978-3-319-57633-6_20 + +Stray, V., Moe, N. B., & Sjoberg, D. I. K. (2020). Daily stand-up meetings: Start breaking the rules. *IEEE Software, 37*(3), 70-77. https://doi.org/10.1109/MS.2018.2875988 + +Stray, V., Sjoberg, D. I. K., & Dyba, T. (2016). The daily stand-up meeting: A grounded theory study. *Journal of Systems and Software, 114*, 101-124. https://doi.org/10.1016/j.jss.2016.01.004 + +Umar, M. A. M. A., Lano, K., & Abubakar, A. K. (2025). Automated requirements engineering framework for agile model-driven development. *Frontiers in Computer Science, 7*, Article 1537100. https://doi.org/10.3389/fcomp.2025.1537100 + +Verwijs, C., & Russo, D. (2024). Do Agile scaling approaches make a difference? An empirical comparison of team effectiveness across popular scaling approaches. *Empirical Software Engineering, 29*, Article 75. https://doi.org/10.1007/s10664-024-10481-5 diff --git a/Aidaily_stage3_prime_rereview_package.md b/Aidaily_stage3_prime_rereview_package.md new file mode 100644 index 0000000..80811dc --- /dev/null +++ b/Aidaily_stage3_prime_rereview_package.md @@ -0,0 +1,72 @@ +# Stage 3' Verification Review Report: Aidaily Protocol Paper + +## Decision + +Minor Revision; proceed to Stage 4.5 Final Integrity after minor cleanup. + +## Field and Reviewer Configuration + +| Role | Configured Identity | Scope | +| --- | --- | --- | +| Field Analyst | Empirical software engineering / HCI protocol analyst | Confirm manuscript genre, revision target, and review checklist. | +| EIC Re-reviewer | Editor for a protocol paper in empirical software engineering and human-centered AI | Verify whether Stage 3 revision requirements were addressed. | +| Editorial Synthesizer | Academic-pipeline synthesis role | Convert verification findings into a pipeline decision. | + +The manuscript is now best characterized as an empirical software engineering and human-centered AI field-study protocol, not a completed results paper. + +## Revision Response Checklist + +### Priority 1 — Required Revisions + +| # | Original Review Comment | Author's Claim | Response Status | Revision Location | Verified? | Quality Assessment | +| --- | --- | --- | --- | --- | --- | --- | +| P1-1 | TTI is under-operationalized; add denominators, sampling windows, coder rules, reliability, adjudication, and examples. | Added TTI eligibility rules, component denominators, examples, stratified double coding, adjudication, and reliability progression criteria. | FULLY_ADDRESSED | Sections 3, 6.5, 9.1, 11, 13.1 | Yes | The revised TTI section now defines event eligibility, sprint windows, component denominators, coding examples, double coding, adjudication, and reliability thresholds. This resolves the reproducibility concern at protocol level. | +| P1-2 | Pilot and confirmatory aims are conflated. | Recast as Stage A feasibility pilot and optional Stage B field evaluation. | FULLY_ADDRESSED | Abstract, Sections 3, 4, 6.5, 13.1-13.2 | Yes | The Stage A/Stage B structure is clear and prevents overclaiming from a small pilot. | +| P1-3 | Intervention behavior is underspecified. | Added intervention reproducibility requirements, prompt taxonomy, confidence bands, escalation rules, writeback constraints, and participant response options. | FULLY_ADDRESSED | Sections 7.1-7.7, 11, 12.3 | Yes | The intervention is now sufficiently specified for a protocol paper. Exact model details remain future implementation fields, but the protocol requires them to be recorded before data collection. | +| P1-4 | AI extraction/linking accuracy is not evaluated. | Added manually coded technical gold sample and performance outcomes for extraction, linking, and conflict detection. | FULLY_ADDRESSED | Sections 2.2, 7.4, 9.2, 11, 13.1-13.2 | Yes | The added gold-sample procedure and technical metrics address the prior blind spot. | +| P1-5 | Ethical safety gates are missing. | Added stop/pause/review rules for psychological safety, workload, prompt burden, opt-out behavior, surveillance complaints, and sensitive-data incidents. | FULLY_ADDRESSED | Sections 3, 13.1, 15.1, 16 | Yes | The safety-gate section is concrete and correctly treats TTI gains as unacceptable if safety thresholds are breached. | +| P1-6 | Participant controls are vague. | Added pause, mark-sensitive, reject, edit, deletion request, appeal, and dashboard restrictions. | FULLY_ADDRESSED | Sections 7.7, 14, 15, 16 | Yes | The control model is now explicit enough for ethics review and implementation planning. | + +### Priority 2 — Suggested Revisions + +| # | Original Review Comment | Response Status | Notes | +| --- | --- | --- | --- | +| P2-7 | Sample-size section lacks actionable assumptions. | FULLY_ADDRESSED | Stage A progression criteria and Stage B power-analysis inputs are now specified. | +| P2-8 | Prompt burden threshold is undefined. | FULLY_ADDRESSED | The revised protocol sets raw NASA-TLX and median prompts-per-day thresholds. | +| P2-9 | Baseline workflow is underspecified. | FULLY_ADDRESSED | Baseline characterization now includes bots, summarizers, Jira automation, issue-linking practices, and dashboards. | +| P2-10 | Literature base is too narrow. | FULLY_ADDRESSED | Traceability, human-centered AI, automation misuse, and workplace monitoring sources were added. | +| P2-11 | Missing-data and partial-consent handling are incomplete. | FULLY_ADDRESSED | Section 13.3 adds rules for partial consent, missing logs, and sensitivity analysis. | + +### Priority 3 — Nice to Fix + +| # | Original Review Comment | Response Status | +| --- | --- | --- | +| P3-12 | Study artifacts are not listed as appendices. | PARTIALLY_ADDRESSED | +| P3-13 | Dissemination is generic. | PARTIALLY_ADDRESSED | +| P3-14 | Governance roles are unclear. | FULLY_ADDRESSED | + +## New Issues Discovered During Re-review + +| # | Type | Location | Description | Severity | +| --- | --- | --- | --- | --- | +| NEW-1 | Protocol artifact completeness | Sections 11, 18 | The protocol names instruments and rubrics, but does not yet provide appendix placeholders for the consent form, survey items, interview guide, TTI coding manual, prompt taxonomy, safety incident form, and data dictionary. | Minor | +| NEW-2 | Reporting/preregistration specificity | Section 17 | The dissemination section says findings will be reported as a protocol-compliant field evaluation, but does not name the preregistration repository, reporting checklist adaptation, or artifact availability plan. | Minor | +| NEW-3 | TTI construct validity | Section 9.1 | The TTI weights are still asserted rather than justified. This is acceptable for a protocol draft if treated as a priori weights, but the final version should state that weights are theory-informed and will not be tuned on outcome data. | Minor | + +## Decision Rationale + +The revised manuscript substantially resolves the Stage 3 major-revision concerns. The most important improvements are the Stage A/Stage B separation, the operational TTI coding rules, the intervention prompt and confidence taxonomy, the technical performance evaluation, and the safety-gate logic. These changes make the protocol auditable and suitable to move into final integrity verification. + +The remaining issues are not major threats to the study design. They are finalization issues: add artifact appendix placeholders, tighten dissemination/preregistration language, and clarify that TTI weights are fixed a priori unless changed through a documented protocol amendment. + +## Residual Minor Revision Actions + +1. Add an appendix roadmap listing planned study artifacts: consent form, survey items, interview guide, TTI coding manual, prompt taxonomy, safety incident form, data dictionary, and analysis code plan. +2. Add one sentence to the TTI section stating that weights are fixed before data collection and any weight change requires a protocol amendment. +3. Expand dissemination by naming the preregistration destination or, if not yet chosen, stating that the registry and artifact repository must be selected before recruitment. + +## Pipeline Decision + +Stage 3' RE-REVIEW result: Minor Revision. + +Per the academic-pipeline state machine, Accept or Minor Revision at Stage 3' proceeds to Stage 4.5 Final Integrity. The residual items can be handled as minor cleanup before or during final integrity preparation; they do not require a Stage 4' major re-revision loop. diff --git a/Aidaily_stage3_review_package.md b/Aidaily_stage3_review_package.md new file mode 100644 index 0000000..f62b4e0 --- /dev/null +++ b/Aidaily_stage3_review_package.md @@ -0,0 +1,213 @@ +# Stage 3 Peer Review Package: Aidaily Protocol Paper + +## Reviewer Configuration + +| Role | Configured Identity | Review Focus | +| --- | --- | --- | +| Editor-in-Chief | Editor for an empirical software engineering / human-centered AI venue | Journal fit, contribution, protocol completeness | +| Reviewer 1 | Mixed-methods and field-experiment methodologist | Design, outcomes, sampling, analysis, reproducibility | +| Reviewer 2 | Agile software engineering and project-management scholar | Literature coverage, domain contribution, Agile/AI framing | +| Reviewer 3 | Human-computer interaction and AI governance reviewer | Practical feasibility, human-AI interaction, ethics, stakeholder impact | +| Devil's Advocate | Adversarial protocol reviewer | Strongest counterarguments, logic gaps, overclaiming | + +## EIC Review Report + +### Overall Recommendation + +Major Revision + +### Confidence Score + +4/5 + +### Summary Assessment + +The manuscript is timely and has a plausible home in empirical software engineering, human-centered AI, or software project management venues. Its strongest feature is the clear positioning of conversational AI as a mediator for team memory rather than as an autonomous decision-maker. The protocol structure is also substantially improved compared with a conceptual-only paper: it includes objectives, hypotheses, participants, intervention, comparator, outcomes, analysis, ethics, and risks. + +The paper is not yet ready as a protocol submission because its intervention and measurement procedures remain underspecified. A reader cannot yet reproduce the AI intervention, compute TTI consistently, audit prompt governance, or judge whether the proposed study is powered for the intended claims. The manuscript states that it is "preregistration-ready" at lines 300-309, but several preregistration-critical fields are still placeholders or policy-level statements. + +### Strengths + +1. **Clear protocol framing:** The paper no longer claims completed results and consistently frames the study prospectively. +2. **Appropriate primary outcome:** TTI is aligned with the intervention's stated purpose and decomposed into meaningful components. +3. **Ethical awareness:** The paper correctly treats psychological safety and surveillance risk as central rather than peripheral. + +### Weaknesses + +1. **Protocol-readiness overstatement:** Lines 300-309 say the protocol is preregistration-ready while listing unresolved design elements. Reframe as "protocol draft" or complete those fields. +2. **Intervention not specified enough:** Lines 129-150 describe functions but not the model, prompt flow, confidence thresholds, human escalation rules, writeback examples, or error handling. +3. **Target venue fit needs sharpening:** The paper should name the intended genre more precisely: protocol paper for empirical software engineering, field study protocol, or HCI intervention protocol. + +## Methodology Review Report + +### Overall Recommendation + +Major Revision + +### Confidence Score + +5/5 + +### Summary Assessment + +The proposed mixed-method, baseline-to-intervention design is appropriate for an early field evaluation, but the protocol lacks the operational precision expected before data collection. The largest methodological risk is that TTI is treated as a primary outcome before its measurement reliability and validity are established. The study can still work, but it should explicitly separate feasibility/pilot aims from confirmatory effectiveness aims. The sample-size section correctly notes clustered outcomes and unknown effect sizes, but it does not define a minimum analyzable unit, power assumptions, or progression criteria. The analysis plan is reasonable at a high level but not yet executable. + +### Strengths + +1. **Design matches the setting:** A quasi-experimental baseline/intervention field design is realistic for workplace software teams. +2. **Mixed-method integration is appropriate:** Joint displays and qualitative explanation are well matched to adoption and governance questions. +3. **Confounds are acknowledged:** Team size, sprint phase, workload intensity, management pressure, and tool maturity are identified. + +### Major Weaknesses + +1. **TTI needs a coding manual:** Lines 160-172 define components, but not denominator construction, sampling windows, coder rules, adjudication, or inter-rater reliability. +2. **Pilot vs confirmatory aims are conflated:** Lines 125-127 say pilot data should estimate effect size, but hypotheses are written as if confirmatory testing is planned. +3. **Power/sample-size plan is incomplete:** The protocol should specify cluster assumptions, expected number of teams, minimum sprint observations, and feasibility thresholds. +4. **Burden threshold is undefined:** H3b at line 66 mentions a "predefined minimal-burden threshold" but no threshold appears later. +5. **Missing-data plan is too general:** Lines 254 and 266 mention missing data and minimization, but there is no rule for missing survey responses, incomplete logs, or opt-out participants within teams. + +### Questions for Authors + +1. Is the first study a feasibility pilot or a powered effectiveness evaluation? +2. What exact rule makes an action item "complete" for CMP? +3. What is the prompt-burden non-inferiority or acceptability threshold? +4. How will teams with partial consent be handled? + +## Domain Review Report + +### Overall Recommendation + +Minor-to-Major Revision + +### Confidence Score + +4/5 + +### Summary Assessment + +The manuscript sits well within current Agile software engineering concerns: distributed coordination, traceability, meeting usefulness, and AI-supported project management. The literature is accurate but thin for a protocol paper. It cites daily stand-up work, Agile scaling, Agile fit, automated requirements engineering, cognitive agents, and AI reporting guidelines. However, it needs stronger grounding in empirical software engineering measurement, traceability, coordination breakdowns, and human-centered AI evaluation. The contribution is plausible: a field protocol for AI-mediated team memory. To make that contribution convincing, the authors should explain how this differs from existing bot-based project-management automation, meeting summarization, and issue-linking tools. + +### Strengths + +1. **Accurate Agile framing:** The paper avoids claiming that Agile ceremonies alone create transparency. +2. **Good distinction between support and decision-making:** Lines 140-146 preserve human authority over project records. +3. **Relevant AI reporting anchor:** The use of CONSORT-AI/SPIRIT-style ideas is appropriate as inspiration, with the non-clinical caveat stated at line 36. + +### Weaknesses + +1. **Missing traceability literature:** The paper should cite software traceability and issue-linking work, not only Agile communication and AI project-management papers. +2. **Limited human-centered AI literature:** The protocol would benefit from references on explainability, automation bias, human oversight, and workplace AI governance. +3. **Competitor/baseline ambiguity:** The comparator is "normal workflow" at lines 152-154, but some teams may already use bots, meeting summarizers, or Jira automation. Baseline tooling must be characterized. + +## Perspective Review Report + +### Overall Recommendation + +Major Revision + +### Confidence Score + +4/5 + +### Summary Assessment + +From an HCI and governance perspective, the protocol asks the right questions but under-specifies the human-AI interaction. The intervention is not just a measurement tool; it changes communication norms in a workplace. That creates risks around consent, opt-out, manager access, social pressure to accept prompts, and chilling effects on surfacing blockers. The manuscript recognizes these risks, but the safeguards remain broad. A strong protocol should include concrete interface states, participant controls, escalation paths, and stopping rules for harm signals. + +### Strengths + +1. **Governance is central:** The paper explicitly rejects individual performance scoring. +2. **Reversibility is included:** Source-linked and reversible writeback is a strong design principle. +3. **Qualitative follow-up is appropriate:** Interviews are necessary to understand trust and chilling effects. + +### Weaknesses + +1. **No participant control model:** The protocol should specify opt-out, pause, delete, redact, and appeal mechanisms. +2. **No harm-monitoring stop rule:** If psychological safety drops or surveillance concerns spike, the protocol should define what happens. +3. **No prompt UX specification:** Prompt types, frequency caps, timeout behavior, and edit/reject flows should be described enough to evaluate burden. +4. **Managerial power asymmetry needs sharper treatment:** Lines 272-280 mention safeguards, but the protocol should separate manager dashboards from team-member views. + +## Devil's Advocate Stress-Test Report + +### Strongest Counter-Argument + +The strongest objection is that the protocol may measure documentation hygiene rather than transparency. A team could improve TTI by creating more linked, complete, and timely records while still becoming less candid in stand-ups or chats because the AI system makes communication feel monitored. In that case, the intervention would optimize the visible artifact layer while degrading the social substrate that Agile communication depends on. The paper partially anticipates this through psychological safety and governance measures, but it does not yet define a decision rule for when improved TTI is outweighed by increased burden, reduced candor, or surveillance concern. Without that rule, the primary outcome could reward a harmful intervention. + +### Issues + +1. **MAJOR: Primary outcome may conflict with ethical outcome.** TTI can rise while psychological safety falls. The protocol needs a composite interpretation rule or safety gate. +2. **MAJOR: AI accuracy is not directly evaluated.** The study records accepted/edit/rejected suggestions, but it does not define gold-standard evaluation for extraction, linking, or conflict detection. +3. **MAJOR: Confirmation can become compliance theater.** Role-aware confirmation may not equal genuine consensus if managers or senior staff are visible in the workflow. +4. **MINOR: The protocol assumes Jira/Git/chat are sufficient operational traces.** Important work may occur in design tools, docs, calls, or private messages. + +### Missing Stakeholder Perspectives + +- Team members who are lower-status or newer to the team. +- Managers who might want individual-level analytics. +- Legal/privacy stakeholders responsible for workplace monitoring. +- Product owners whose scope decisions may be challenged by AI-generated traceability. + +## Editorial Synthesis + +### Reviewer Summary Matrix + +| Reviewer | Recommendation | Confidence | Main Concern | +| --- | --- | --- | --- | +| EIC | Major Revision | 4 | Protocol-readiness and intervention specification | +| R1 Methodology | Major Revision | 5 | TTI operationalization, pilot/confirmatory ambiguity, power plan | +| R2 Domain | Minor-to-Major Revision | 4 | Thin traceability/HCAI literature and baseline characterization | +| R3 Perspective | Major Revision | 4 | Human-AI interaction and governance safeguards | +| Devil's Advocate | Major issues | n/a | TTI may improve while candor and psychological safety decline | + +### Editorial Decision + +Major Revision + +The manuscript is promising and substantially improved, but not ready for acceptance as a protocol paper. The required changes are feasible and do not require abandoning the study. The central revision task is to make the protocol executable: define the intervention, measurement rules, safety gates, sample/power logic, and human-AI governance model in enough detail that another research team could implement or audit the study. + +## Revision Roadmap + +### P1: Must Fix + +| # | Issue | Section | Required Action | +| --- | --- | --- | --- | +| 1 | TTI is under-operationalized. | Outcomes / Analysis | Add a TTI coding manual: denominators, sampling windows, coder training, inter-rater reliability, adjudication, examples. | +| 2 | Pilot and confirmatory aims are conflated. | Objectives / Design / Sample Size | State whether this is a feasibility pilot, confirmatory study, or two-phase program. Align hypotheses and analyses accordingly. | +| 3 | Intervention behavior is underspecified. | Intervention | Add model/system description, prompt taxonomy, confidence thresholds, escalation rules, writeback examples, failure modes, and error handling. | +| 4 | AI extraction/linking accuracy is not evaluated. | Outcomes / Analysis | Add technical performance outcomes: precision/recall or agreement against a manually coded gold sample for extraction, linking, and conflict detection. | +| 5 | Ethical safety gates are missing. | Ethics / Risk Management | Define stop/pause/review rules for psychological safety decline, high prompt burden, opt-out concerns, or surveillance complaints. | +| 6 | Participant control mechanisms are vague. | Ethics / Intervention | Specify opt-out, pause, redact, edit, reject, delete, and appeal mechanisms. | + +### P2: Should Fix + +| # | Issue | Section | Suggested Action | +| --- | --- | --- | --- | +| 7 | Sample-size section lacks actionable assumptions. | Participants | Add pilot progression criteria and, for confirmatory study, ICC/effect-size assumptions or a simulation-based power plan. | +| 8 | Prompt burden threshold is undefined. | Hypotheses / Outcomes | Define the non-inferiority or acceptability threshold for workload/prompt burden. | +| 9 | Baseline workflow is underspecified. | Comparator | Characterize existing automation, meeting summarizers, bots, and issue-linking practices. | +| 10 | Literature base is too narrow. | Introduction | Add software traceability, human-centered AI, automation bias, workplace monitoring, and team cognition references. | +| 11 | Missing-data and partial-consent handling are incomplete. | Data Management / Analysis | Add rules for incomplete logs, survey nonresponse, and teams where not everyone consents. | + +### P3: Consider + +| # | Issue | Section | Optional Improvement | +| --- | --- | --- | --- | +| 12 | Study artifacts are not listed. | Appendix | Add planned appendices: consent form, survey items, interview guide, TTI rubric, prompt taxonomy. | +| 13 | Dissemination is generic. | Dissemination | Specify reporting checklist and preregistration repository. | +| 14 | Governance roles are unclear. | Ethics | Add data steward, technical owner, team representative, and escalation owner. | + +## Required Author Response + +The revision should include a response table with one row per P1/P2 issue: + +| Concern ID | Reviewer Source | Action Taken | Manuscript Location | Status | +| --- | --- | --- | --- | --- | +| P1-1 | R1 / EIC | | | | +| P1-2 | R1 | | | | +| P1-3 | EIC / R3 | | | | +| P1-4 | DA / R1 | | | | +| P1-5 | R3 / DA | | | | +| P1-6 | R3 | | | | + +## Pipeline Decision + +Stage 3 REVIEW result: Major Revision. Proceed to Stage 4 REVISE if the user confirms. diff --git a/Aidaily_stage4_5_final_integrity_report.md b/Aidaily_stage4_5_final_integrity_report.md new file mode 100644 index 0000000..619c537 --- /dev/null +++ b/Aidaily_stage4_5_final_integrity_report.md @@ -0,0 +1,74 @@ +# Stage 4.5 Final Integrity Report: Aidaily Protocol Paper + +## Verdict + +PASS. + +The revised protocol has no detected fabricated references, dangling citations, orphan references, completed-results overclaiming, or unresolved Stage 3' minor-revision blockers. + +## Materials Checked + +- Manuscript: `Aidaily_protocol_paper.md` +- Re-review package: `Aidaily_stage3_prime_rereview_package.md` +- Stage 4 revision package: `Aidaily_stage4_revision_package.md` + +## Local Consistency Checks + +| Check | Result | +| --- | --- | +| Word count | 6,181 words | +| Heading sequence | Material Passport, Abstract, Sections 1-20, References | +| Completed-results language | No flagged phrases found | +| Stage 3' minor cleanups | Completed | +| Reference list vs in-text citations | All listed references are cited in text; all in-text citations have reference-list entries | +| Protocol status | Correctly states draft is not preregistered, ethics-approved, or implementation-ready | + +## Reference Verification Audit Trail + +| Reference | Query Used | Verification Source | Verdict | +| --- | --- | --- | --- | +| Ball (2021) | `Ball 2021 Electronic monitoring and surveillance in the workplace doi 10.2760/451453` | European Commission JRC repository: https://publications.jrc.ec.europa.eu/repository/handle/JRC125716 | VERIFIED | +| Chan et al. (2013) | `Chan Tetzlaff Altman Laupacis SPIRIT 2013 Annals Internal Medicine 158 3 200 207 DOI` | PubMed / Annals of Internal Medicine: https://pubmed.ncbi.nlm.nih.gov/23295957/ | VERIFIED | +| Cinkusz et al. (2025) | `Cinkusz Chudziak Niewiadomska-Szynkiewicz 2025 Cognitive agents powered by large language models for agile software project management Electronics 14 87` | MDPI Electronics: https://www.mdpi.com/2079-9292/14/1/87 | VERIFIED | +| Cleland-Huang et al. (2014) | `Cleland-Huang Gotel Huffman Hayes Mäder Zisman 2014 Software traceability trends and future directions DOI` | ACM Digital Library: https://dl.acm.org/doi/10.1145/2593882.2593891 | VERIFIED | +| Edmondson (1999) | `"Psychological safety and learning behavior in work teams" "Administrative Science Quarterly" "350" "383"` | SAGE Journals: https://journals.sagepub.com/doi/10.2307/2666999 | VERIFIED | +| Hart (2006) | `Hart 2006 NASA Task Load Index 20 years later Proceedings Human Factors Ergonomics Society 50 9 904 908 DOI` | SAGE Journals: https://journals.sagepub.com/doi/10.1177/154193120605000909 | VERIFIED | +| Itzik & Roy (2023) | `Itzik Roy 2023 Does agile methodology fit all characteristics of software projects Empirical Software Engineering 28 105 DOI` | ACM / Springer DOI record: https://dl.acm.org/doi/10.1007/s10664-023-10334-7 | VERIFIED | +| Liu et al. (2020) | `Liu Cruz Rivera Moher Calvert Denniston SPIRIT-AI CONSORT-AI 2020 Nature Medicine 26 1364 1374 DOI` | PubMed / Nature Medicine: https://pubmed.ncbi.nlm.nih.gov/32908283/ | VERIFIED | +| Malla (2025) | `Malla 2025 Analyzing the impact of agile methodologies on software quality and delivery speed comparative study DOI 10.30574/wjarr.2025.25.1.0184` | WJARR journal page: https://wjarr.com/content/analyzing-impact-agile-methodologies-software-quality-and-delivery-speed-comparative-study | VERIFIED | +| Parasuraman & Riley (1997) | `Parasuraman Riley 1997 Humans and automation use misuse disuse abuse Human Factors 39 2 230 253 DOI` | SAGE Journals: https://journals.sagepub.com/doi/10.1518/001872097778543886 | VERIFIED | +| Shneiderman (2020) | `Shneiderman 2020 Human-centered artificial intelligence reliable safe trustworthy International Journal Human-Computer Interaction 36 6 495 504 DOI` | Taylor & Francis: https://www.tandfonline.com/doi/full/10.1080/10447318.2020.1741118 | VERIFIED | +| Stray et al. (2017) | `Stray Moe Bergersen 2017 Are daily stand-up meetings valuable survey developers software teams DOI 10.1007/978-3-319-57633-6_20` | Springer: https://link.springer.com/chapter/10.1007/978-3-319-57633-6_20 | VERIFIED | +| Stray et al. (2020) | `Stray Moe Sjoberg 2020 Daily stand-up meetings Start breaking the rules IEEE Software 37 3 70 77 DOI` | IEEE Xplore: https://ieeexplore.ieee.org/document/8501962/ | VERIFIED | +| Stray et al. (2016) | `Stray Sjoberg Dyba 2016 The daily stand-up meeting grounded theory study Journal of Systems and Software 114 101 124 DOI` | Journal record / DOI result: https://doi.org/10.1016/j.jss.2016.01.004 | VERIFIED | +| Umar et al. (2025) | `Umar Lano Abubakar 2025 Automated requirements engineering framework agile model-driven development Frontiers in Computer Science 7 1537100 DOI` | Frontiers in Computer Science: https://www.frontiersin.org/journals/computer-science/articles/10.3389/fcomp.2025.1537100/full | VERIFIED | +| Verwijs & Russo (2024) | `Verwijs Russo 2024 Do Agile scaling approaches make a difference empirical comparison team effectiveness Empirical Software Engineering 29 75 DOI` | Springer Empirical Software Engineering: https://link.springer.com/article/10.1007/s10664-024-10481-5 | VERIFIED | + +## Citation Context Check + +| Citation Cluster | Manuscript Claim | Integrity Assessment | +| --- | --- | --- | +| Stray et al. (2016, 2017, 2020) | Daily stand-ups support awareness and coordination but vary in value and should be adapted to team needs. | Supported by cited stand-up literature. | +| Verwijs & Russo (2024); Itzik & Roy (2023) | Agile scaling/framework fit is context-sensitive and not reducible to framework compliance. | Supported. | +| Umar et al. (2025); Cinkusz et al. (2025); Malla (2025) | AI and automation can support requirements/project-management workflows while raising workflow-fit and over-reliance concerns. | Supported, with the manuscript using cautious language. | +| Cleland-Huang et al. (2014) | Traceability is valuable but often ad hoc and after the fact. | Supported. | +| Shneiderman (2020); Parasuraman & Riley (1997); Ball (2021) | Human-centered AI, automation misuse, and workplace monitoring risks motivate human control and governance safeguards. | Supported. | +| Hart (2006); Edmondson (1999) | NASA-TLX and psychological safety are appropriate constructs/instruments for workload and team climate. | Supported. | +| Chan et al. (2013); Liu et al. (2020) | Protocol completeness and AI-intervention reporting principles inform the protocol structure. | Supported, with explicit caveat that the study is not a clinical trial. | + +## Corrections Applied Before PASS + +1. Added fixed-a-priori language for TTI weights and protocol amendments. +2. Expanded dissemination language to require preregistration and artifact-repository selection before recruitment. +3. Added an appendix roadmap for consent, data dictionary, TTI coding manual, prompt examples, survey items, interview guide, gold-sample guide, safety incident workflow, and analysis code plan. +4. Updated Material Passport status to `STAGE_4_5_FINAL_INTEGRITY_PASS`. + +## Residual Notes + +No blocking integrity issues remain. Before real-world recruitment, the study team must still complete ethics review, preregistration, actual instrument text, consent language, data-processing agreements, and organization-specific data access controls. + +## Pipeline Decision + +Stage 4.5 FINAL INTEGRITY result: PASS. + +Per the academic-pipeline state machine, the manuscript is ready for Stage 5 FINALIZE. diff --git a/Aidaily_stage4_revision_package.md b/Aidaily_stage4_revision_package.md new file mode 100644 index 0000000..293de74 --- /dev/null +++ b/Aidaily_stage4_revision_package.md @@ -0,0 +1,39 @@ +# Stage 4 Revision Package: Aidaily Protocol Paper + +## Revision Target + +- Revised manuscript: `Aidaily_protocol_paper.md` +- Prior review package: `Aidaily_stage3_review_package.md` +- Stage 3 decision: Major Revision +- Stage 4 status: Revised and ready for re-review + +## Major Changes Completed + +1. Recast the design as a two-stage program: Stage A feasibility pilot and optional Stage B powered field evaluation. +2. Operationalized TTI with eligibility rules, denominator construction, sprint-level sampling windows, examples, coder reliability checks, adjudication, and coding-manual freeze criteria. +3. Specified the AI intervention through model/version logging, prompt taxonomy, confidence bands, escalation rules, writeback limits, and participant response options. +4. Added technical performance outcomes for extraction, artifact linking, and conflict detection against a manually coded gold sample. +5. Added safety gates for psychological safety, workload, prompt burden, opt-out behavior, manager misuse, and sensitive-data incidents. +6. Added participant controls: pause, mark sensitive, reject, edit, delete request, and appeal. +7. Added missing-data and partial-consent handling rules. +8. Expanded the literature base with software traceability, human-centered AI, automation misuse, and workplace monitoring sources. +9. Reframed protocol status as a revised draft for preregistration and ethics review, not as preregistration-ready or approved. + +## Author Response Location + +The full response matrix is embedded in `Aidaily_protocol_paper.md` under: + +- `## 19. Stage 4 Revision Response Matrix` + +It covers all P1 and P2 concerns from Stage 3. + +## Verification Checks + +- Word count: 6,008 words. +- Completed-results language check: no flagged phrases found for "results show", "we found", "was conducted", "participants reported", "significant improvement", "improved by", "reduced by", "findings show", or "data showed". +- Metadata updated: `Verification Status: STAGE_4_REVISED_UNVERIFIED`; `Version Label: aidaily_protocol_paper_v2_stage4`. +- Added references are cited in text and listed in References. + +## Recommended Next Step + +Run Stage 3 re-review on `Aidaily_protocol_paper.md`, using the embedded response matrix as the author response. diff --git a/Aidaily_stage5_finalization_package.md b/Aidaily_stage5_finalization_package.md new file mode 100644 index 0000000..46c5798 --- /dev/null +++ b/Aidaily_stage5_finalization_package.md @@ -0,0 +1,40 @@ +# Stage 5 Finalization Package: Aidaily Protocol Paper + +## Finalization Verdict + +Stage 5 FINALIZE is complete for Markdown, DOCX, HTML, and RTF outputs. PDF generation is blocked in this workspace because Pandoc, LaTeX, wkhtmltopdf, LibreOffice, and a usable CUPS HTML/RTF-to-PDF filter are not installed. + +## Final Manuscript Files + +| File | Purpose | Notes | +| --- | --- | --- | +| `Aidaily_final_manuscript.md` | Clean submission manuscript | Internal Material Passport and Stage 4 response matrix removed. | +| `Aidaily_final_manuscript.docx` | Word-compatible export | Generated with macOS `textutil`; use for review, not as a fully journal-styled DOCX. | +| `Aidaily_final_manuscript.html` | HTML export | Generated with macOS `textutil`. | +| `Aidaily_final_manuscript.rtf` | Rich-text export | Generated with macOS `textutil`. | + +## Preserved Audit Files + +| File | Purpose | +| --- | --- | +| `Aidaily_protocol_paper.md` | Integrity-passed full working manuscript with Material Passport and response matrix. | +| `Aidaily_stage3_review_package.md` | Initial Stage 3 peer review. | +| `Aidaily_stage4_revision_package.md` | Revision summary. | +| `Aidaily_stage3_prime_rereview_package.md` | Stage 3' verification review. | +| `Aidaily_stage4_5_final_integrity_report.md` | Final integrity audit trail and PASS verdict. | + +## Checks Performed + +| Check | Result | +| --- | --- | +| Clean manuscript word count | 5,774 words | +| Internal pipeline metadata removed from clean manuscript | PASS | +| Stage 4 response matrix removed from clean manuscript | PASS | +| References retained in clean manuscript | PASS | +| PDF conversion attempted | BLOCKED: local converter unavailable | + +## Pipeline Decision + +Stage 5 FINALIZE result: COMPLETE WITH PDF BLOCKED BY LOCAL TOOLING. + +The manuscript is ready for Stage 6 PROCESS SUMMARY. diff --git a/Aidaily_stage6_process_record.md b/Aidaily_stage6_process_record.md new file mode 100644 index 0000000..74a91ce --- /dev/null +++ b/Aidaily_stage6_process_record.md @@ -0,0 +1,125 @@ +# Aidaily Paper Creation Process Record + +## Project + +- Working title: Protocol for Evaluating a Conversational AI Framework for Agile Team Transparency and Knowledge Traceability +- Source material: `.context/attachments/FIk5qr/Aidaily v0.2 (1).pdf` +- Workspace: `/Users/trovo/conductor/workspaces/Starter project/chengdu` +- Pipeline: Academic Research Suite full academic pipeline +- Process date: 2026-06-26 +- Final pipeline status: Complete, with PDF blocked by unavailable local conversion tooling + +## Final Outputs + +| Artifact | Description | Status | +| --- | --- | --- | +| `Aidaily_final_manuscript.md` | Clean submission manuscript | Complete | +| `Aidaily_final_manuscript.docx` | Word-compatible export generated with macOS `textutil` | Complete | +| `Aidaily_final_manuscript.html` | HTML export generated with macOS `textutil` | Complete | +| `Aidaily_final_manuscript.rtf` | Rich-text export generated with macOS `textutil` | Complete | +| `Aidaily_protocol_paper.md` | Integrity-passed working manuscript with Material Passport and response matrix | Complete | +| `Aidaily_stage4_5_final_integrity_report.md` | Final reference and citation audit trail | PASS | +| PDF output | PDF manuscript | Blocked: no Pandoc, LaTeX, wkhtmltopdf, LibreOffice, or usable CUPS HTML/RTF-to-PDF filter | + +## Pipeline Timeline + +| Stage | Action | Output | +| --- | --- | --- | +| Intake | Read the attached PDF manuscript and extracted its text. | Identified the draft as a short AI/Agile manuscript with unsupported completed-results language and no formal References section. | +| Stage 2.5 Integrity | Checked the existing draft before revision. | Integrity failed because the draft described completed evaluation results without evidence and lacked verifiable references. | +| Revision to conceptual/protocol direction | Converted the manuscript away from unsupported empirical-result claims. | `Aidaily_v0.3_revised.md` | +| Protocol-paper drafting | Reframed the work as a prospective mixed-method human study protocol. | `Aidaily_protocol_paper.md` initial protocol version | +| Stage 3 Review | Simulated multi-perspective peer review. | `Aidaily_stage3_review_package.md`; decision: Major Revision | +| Stage 4 Revision | Addressed all P1/P2 review concerns. | Added Stage A/Stage B design, TTI operationalization, prompt taxonomy, safety gates, participant controls, technical performance outcomes, and missing-data rules. | +| Stage 3' Re-review | Verified whether revision addressed reviewer concerns. | `Aidaily_stage3_prime_rereview_package.md`; decision: Minor Revision | +| Minor cleanup | Resolved the three residual re-review issues. | Added fixed-a-priori TTI weights, preregistration/artifact repository language, and appendix roadmap. | +| Stage 4.5 Final Integrity | Verified citations, reference list, citation context, protocol status, and overclaiming risk. | `Aidaily_stage4_5_final_integrity_report.md`; verdict: PASS | +| Stage 5 Finalize | Created clean final manuscript outputs. | Markdown, DOCX, HTML, and RTF outputs complete; PDF blocked by local tooling. | +| Stage 6 Process Summary | Generated this process record. | `Aidaily_stage6_process_record.md` | + +## Major Content Decisions + +1. The original draft was not treated as a completed empirical study because it did not contain verifiable methods, data, or results. +2. The paper was reframed as a protocol paper to preserve the research idea while removing unsupported outcome claims. +3. The primary study design was split into Stage A feasibility and optional Stage B field evaluation to avoid overclaiming from a small pilot. +4. The Team Transparency Index was retained but made operational through sprint-level coding windows, component denominators, event eligibility rules, coder reliability checks, and fixed a priori weights. +5. The AI intervention was specified as a human-confirmed mediator rather than an autonomous project-management actor. +6. Safety and governance were treated as hard constraints, not secondary discussion points: TTI gains are unacceptable if psychological safety, workload, prompt burden, opt-out, surveillance, or sensitive-data thresholds are breached. +7. The final clean manuscript excludes internal pipeline metadata and reviewer response tables, while the audit manuscript preserves them for traceability. + +## Review and Revision Summary + +The Stage 3 review identified six required issues: + +1. Under-operationalized TTI. +2. Conflated feasibility and confirmatory aims. +3. Underspecified AI intervention behavior. +4. Missing AI extraction/linking/conflict-detection accuracy outcomes. +5. Missing ethical safety gates. +6. Vague participant control mechanisms. + +All six were fully addressed by Stage 4 and verified in Stage 3' re-review. + +The Stage 3' re-review left three minor issues: + +1. Add appendix roadmap. +2. Clarify TTI weights are fixed a priori. +3. Tighten preregistration and dissemination language. + +All three were applied before final integrity verification. + +## Integrity Summary + +Final integrity status: PASS. + +The final integrity audit found: + +- No fabricated or unverified references. +- No dangling in-text citations. +- No orphan references. +- No flagged completed-results language. +- No remaining claim that the protocol is preregistered, ethics-approved, or implementation-ready. +- All citation-context checks were consistent with the cautious claims made in the manuscript. + +Reference verification sources included: + +- European Commission Joint Research Centre. +- PubMed. +- MDPI. +- ACM Digital Library. +- SAGE Journals. +- Springer. +- Nature Medicine. +- WJARR. +- Taylor & Francis. +- IEEE Xplore. +- Frontiers. + +## File Inventory + +| File | Role | +| --- | --- | +| `Aidaily_v0.3_revised.md` | Revised conceptual/protocol bridge draft | +| `Aidaily_protocol_paper.md` | Integrity-passed working manuscript | +| `Aidaily_stage3_review_package.md` | Initial peer-review simulation | +| `Aidaily_stage4_revision_package.md` | Revision handoff package | +| `Aidaily_stage3_prime_rereview_package.md` | Verification re-review package | +| `Aidaily_stage4_5_final_integrity_report.md` | Final integrity report | +| `Aidaily_final_manuscript.md` | Clean final manuscript | +| `Aidaily_final_manuscript.docx` | Word-compatible final manuscript | +| `Aidaily_final_manuscript.html` | HTML final manuscript | +| `Aidaily_final_manuscript.rtf` | Rich-text final manuscript | +| `Aidaily_stage5_finalization_package.md` | Finalization record | +| `Aidaily_stage6_process_record.md` | Process record | + +## Remaining Practical Tasks Before Real Submission + +1. Choose a target venue and adapt formatting to that venue's author instructions. +2. Complete ethics review, consent language, organization-specific data-processing agreements, and access-control plan before recruitment. +3. Finalize the actual survey instruments, interview guide, TTI coding manual, prompt examples, safety incident form, and analysis plan. +4. Select the preregistration destination and artifact repository before recruitment. +5. Generate a PDF after installing a converter such as Pandoc with LaTeX, wkhtmltopdf, or LibreOffice. + +## Process Result + +The ARS full pipeline has produced a clean, integrity-checked protocol manuscript and a complete audit trail. The only incomplete output format is PDF, blocked by missing local conversion tooling rather than manuscript content. diff --git a/Aidaily_v0.3_revised.md b/Aidaily_v0.3_revised.md new file mode 100644 index 0000000..384b0cb --- /dev/null +++ b/Aidaily_v0.3_revised.md @@ -0,0 +1,205 @@ +# Conversational AI for Agile Transparency: A Conceptual Framework for Team Alignment, Knowledge Traceability, and Human-AI Collaboration + +## Abstract + +Agile software development depends on frequent communication, shared context, and fast correction of misunderstandings. Daily stand-ups, chat threads, issue trackers, and version control systems all contain partial traces of team activity, but these traces are often fragmented across tools and are rarely reconciled into a durable record of decisions, blockers, and action commitments. This paper proposes a conversational AI framework for improving transparency in Agile teams by connecting unstructured team communication with structured project artifacts such as Jira issues and Git activity. The framework uses natural language processing, traceability linking, conflict detection, and role-aware consensus prompts to convert informal discussion into validated knowledge artifacts. It also defines a Team Transparency Index (TTI), a composite metric that combines coverage, consistency, consensus, timeliness, and completeness. Because the present paper is a conceptual framework and evaluation protocol, it does not claim empirical performance gains. Instead, it specifies testable hypotheses, operational measures, data sources, and study procedures for future field evaluation. The contribution is a sociotechnical model in which AI supports team memory and documentation without replacing human judgment. + +**Keywords:** Agile software development; conversational AI; daily stand-ups; knowledge management; traceability; large language models; team transparency; human-AI collaboration + +## 1. Introduction + +Agile methods place communication at the center of software development. Short feedback loops, daily coordination, adaptive planning, and collective ownership depend on the ability of team members to maintain a shared view of what has been decided, what remains blocked, and what work is actually moving through the delivery system. In practice, however, this shared view is difficult to sustain. A team may discuss a blocker in a stand-up, resolve part of it in a chat thread, record a partial update in Jira, and complete the implementation through commits and pull requests. Each channel is useful, but no single channel reliably preserves the full meaning of the work. + +Daily stand-up meetings illustrate the problem. Prior research shows that stand-ups are widely used and can improve awareness, but their value varies by team context, role, and meeting quality (Stray et al., 2017). Other work on daily stand-ups emphasizes that the practice can become burdensome or ineffective when it is treated as a ritual rather than a communication mechanism (Stray et al., 2020). These findings suggest that Agile transparency is not guaranteed by ceremony adoption alone. Teams need mechanisms that preserve useful coordination signals while reducing avoidable meeting and documentation overhead. + +The same issue appears in scaled or distributed Agile environments. Research on Agile scaling suggests that the selected scaling framework is less important than the team's actual effectiveness, autonomy, stakeholder alignment, and management context (Verwijs & Russo, 2024). Work on Agile methodology fit also shows that Agile practices must be adapted to project characteristics rather than applied uniformly (Itzik & Roy, 2023). These findings motivate a transparency layer that is adaptive, evidence-linked, and sensitive to team context. + +Recent advances in large language models (LLMs), automated requirements engineering, and cognitive agent systems create a new opportunity. AI systems can summarize unstructured text, identify entities and commitments, connect artifacts across tools, and generate targeted clarification prompts. Automated requirements engineering research shows how machine learning can extract structured models from natural language requirements in Agile model-driven development (Umar et al., 2025). LLM-based multi-agent work such as CogniSim explores how AI agents can simulate or support Agile project roles and workflows (Cinkusz et al., 2025). These systems point toward a broader use of AI in software project management, but many applications still focus on automation, prediction, or simulated task performance rather than the human communication layer. + +This paper proposes a conversational AI framework that acts as a team knowledge mediator. The system listens to communication artifacts, links them to project records, detects inconsistency, and asks the relevant people to confirm or correct interpretations. The goal is not to make project decisions automatically. The goal is to preserve and verify the team's own decisions so that team memory becomes traceable, reversible, and easier to query. + +The paper addresses three research questions: + +1. How can conversational AI transform informal Agile communication into validated and traceable knowledge artifacts? +2. Which measurable dimensions can operationalize team transparency in a way that supports empirical evaluation? +3. What evaluation protocol can test whether such a framework improves alignment without increasing cognitive burden or reducing psychological safety? + +## 2. Related Work + +### 2.1 Daily Stand-Ups and Agile Communication + +Daily stand-ups are one of the most recognizable Agile practices. Stray et al. (2017) found that the practice was common among Agile teams, but that perceived value varied across developers and team settings. Junior developers tended to report more positive perceptions, while senior developers and members of larger teams were more skeptical. This variation matters because it suggests that stand-ups do not automatically produce shared understanding. Their effect depends on relevance, team size, facilitation, and whether information raised in the meeting leads to useful follow-up. + +Earlier grounded theory work also found that daily stand-ups support coordination through awareness, problem solving, and shared monitoring, while still being sensitive to meeting structure and team context (Stray et al., 2016). Stray et al. (2020) further argued that teams may need to adapt or break conventional stand-up rules when the ritual does not serve team needs. The implication for AI support is clear: a conversational agent should not simply record every utterance or force every update into a rigid template. It should detect what is actionable, uncertain, contradicted, or underdocumented, and then ask for clarification only when the expected benefit justifies the interruption. + +### 2.2 Agile Transparency, Context Fit, and Scaling + +Transparency is often discussed as a principle of Agile work, but it is difficult to measure. Verwijs and Russo (2024) compared team effectiveness across Agile scaling approaches and found that differences among scaling frameworks were minor in practical terms. This shifts attention from framework labels to local conditions such as autonomy, stakeholder satisfaction, responsiveness, and team-level coordination. Similarly, Itzik and Roy (2023) argue that Agile methodology fit depends on project characteristics and should be evaluated through a decision framework rather than assumed universally. + +These studies support the need for adaptive transparency tools. A lightweight team working on a small product may require minimal AI mediation, while a distributed team with multiple handoffs may need stronger traceability and confirmation rules. The framework proposed here treats transparency as a configurable sociotechnical capability rather than a fixed reporting practice. + +### 2.3 AI in Agile Project Management + +AI support for Agile work has expanded across requirements engineering, project analytics, and agent-based project management. Umar et al. (2025) proposed an automated requirements engineering framework that uses machine learning to extract structured class-diagram components from textual requirements. Their work demonstrates that natural language artifacts in Agile settings can be formalized into more structured representations. + +Cinkusz et al. (2025) proposed CogniSim, a cognitive multi-agent system powered by LLMs for Agile software project management. The framework uses virtual agents to represent roles such as product owner, architect, and QA engineer, and evaluates their ability to support project workflows in simulated environments. This line of work shows that LLMs can support complex project management functions, but it also raises questions about where automation should stop. In real teams, the legitimacy of decisions depends on human accountability, shared context, and role-based agreement. + +Malla (2025) compares Agile methods and technology-enhanced practices, including AI-enabled tools and hybrid Agile-Kanban workflows. The reported findings suggest potential gains in delivery speed and quality when advanced tools are integrated carefully, but also note risks such as over-reliance on automation. This warning is central to the present framework: conversational AI should support human oversight by making claims traceable and confirmations explicit. + +### 2.4 Research Gap + +Existing research has addressed Agile communication, Agile framework selection, automated requirements engineering, and AI-supported project management. Less attention has been given to the specific problem of continuous team memory: how informal decisions and action commitments move from conversation into durable, verified project knowledge. The proposed framework fills this gap by focusing on sociotechnical alignment rather than only prediction, task automation, or simulated agent performance. + +## 3. Conceptual Framework + +### 3.1 Overview + +The proposed system is a conversational AI mediator embedded in the team's communication environment. It connects three categories of input: + +1. **Team communication:** daily stand-up transcripts, asynchronous chat messages, issue comments, and decision threads. +2. **Project management data:** Jira issue status, assignees, sprint membership, priorities, due dates, and issue transitions. +3. **Version control data:** commits, pull requests, branch names, review comments, merge events, and references to issue identifiers. + +The system converts these inputs into structured candidate knowledge items, links them to system-of-record artifacts, checks for inconsistency, and prompts the team for confirmation when confidence is insufficient or when the decision has meaningful impact. + +### 3.2 Ingestion Layer + +The ingestion layer collects and normalizes data from communication and development tools. Speech-to-text modules may be used for synchronous meetings. Webhooks and APIs can be used for Jira, Git hosting platforms, and chat systems such as Rocket.Chat or Slack. The layer normalizes time zones, user identifiers, project identifiers, and artifact references so that later modules can compare events across tools. + +The ingestion layer should preserve provenance. Every extracted claim or action item must retain a pointer to its source message, issue, commit, or transcript segment. Provenance is necessary for explainability, correction, and auditability. + +### 3.3 NLP and Traceability Pipeline + +The NLP pipeline transforms communication into candidate events. It performs five functions. + +First, segmentation divides transcripts and messages into speaker-attributed utterances. Second, entity and relation extraction identifies people, tasks, artifacts, dates, blockers, and status claims. Third, action and decision mining extracts commitments such as "I will refactor the payment module tomorrow" or decisions such as "We will postpone the analytics dashboard until the next sprint." Fourth, traceability linking maps extracted items to Jira issues, Git commits, pull requests, or repository components using explicit identifiers and semantic similarity. Fifth, conflict detection compares claims across tools, for example when a developer says a task is done while the corresponding issue remains in progress. + +The pipeline should produce candidate records rather than final facts. Each record includes extracted content, confidence, source evidence, linked artifacts, and a suggested verification pathway. + +### 3.4 Consensus and Alignment Engine + +The consensus engine determines when and how the system should ask the team for confirmation. Low-risk updates, such as adding a summary comment to a Jira issue, may require only the assignee's confirmation. Higher-impact decisions, such as changing scope or redefining acceptance criteria, may require confirmation from multiple roles such as developer, QA, product owner, and scrum master. + +The engine applies three principles: + +1. **Minimal interruption:** ask only when uncertainty, inconsistency, or impact warrants attention. +2. **Role-aware confirmation:** request input from people who have responsibility or context for the item. +3. **Reversibility:** every AI-proposed update must be visible, attributable, and reversible. + +Consensus is therefore not a vote on truth in the abstract. It is an operational rule for deciding whether the team has sufficiently validated a proposed knowledge artifact. + +### 3.5 Knowledge Store and Writeback + +Confirmed items are stored in a team knowledge graph or vector-backed knowledge store. Entities may include tasks, components, blockers, decisions, action items, people, commits, and meetings. Relationships may include "blocks," "implements," "duplicates," "depends on," "decided in," and "confirmed by." + +The writeback module synchronizes confirmed items to the relevant system of record. Examples include adding a Jira comment, updating an issue status, generating an architecture decision record, summarizing a resolved blocker, or suggesting a pull request description. The system should not silently overwrite authoritative fields. For fields that affect scope, dates, ownership, or acceptance criteria, writeback should require explicit approval. + +## 4. Team Transparency Index + +The Team Transparency Index (TTI) is a proposed composite measure for evaluating whether the system improves shared understanding and traceability. It is not intended as a universal productivity score. It measures the quality of alignment among communication, project records, and team-confirmed knowledge. + +The index contains five normalized components: + +| Component | Description | +| --- | --- | +| Coverage (COV) | Ratio of communication-mentioned tasks or decisions that are linked to Jira, Git, or another project artifact. | +| Consistency (CON) | Degree of factual alignment between communication claims and structured project records. | +| Consensus (CSN) | Proportion of decisions or high-impact updates that achieve the required role-based confirmation threshold. | +| Timeliness (TML) | Speed with which confirmed updates appear in the appropriate system of record, normalized so shorter delays score higher. | +| Completeness (CMP) | Share of action items containing who, what, and when metadata. | + +The proposed formula is: + +```text +TTI = 0.25*COV + 0.25*CON + 0.20*CSN + 0.15*TML + 0.15*CMP +``` + +The weights reflect an initial design assumption: coverage and consistency are foundational, consensus is critical for legitimacy, and timeliness and completeness improve operational usefulness. Future empirical work should test the sensitivity of these weights and may adjust them by team type, domain, or project risk. + +## 5. Evaluation Protocol + +Because the current manuscript presents a framework rather than completed empirical results, this section defines a study design for future validation. + +### 5.1 Study Design + +A mixed-method field study is recommended. Participating Agile teams would use the conversational AI framework during multiple sprints, with a baseline period before deployment and an intervention period after deployment. A comparison group or staggered rollout would strengthen causal inference if feasible. + +### 5.2 Participants and Setting + +The target setting is a software development organization using issue tracking, Git-based version control, and synchronous or asynchronous stand-up communication. A pilot study could begin with one Scrum team of approximately 6 to 10 members, followed by a broader multi-team evaluation. Participant roles should include developers, QA engineers, product owners, and scrum masters or team leads. + +### 5.3 Measures + +Quantitative measures should include TTI and its five components, mean time to detect inconsistencies, number of unresolved communication-to-record conflicts, documentation completeness, and clarification prompt frequency. Team-level survey measures should assess perceived transparency, cognitive burden, trust in AI suggestions, and psychological safety. + +Qualitative data should include semi-structured interviews, observation notes, and examples of accepted, rejected, and corrected AI suggestions. These data are necessary because a high TTI score may still be harmful if it is achieved through intrusive prompts or surveillance-like behavior. + +### 5.4 Hypotheses + +The evaluation should test the following hypotheses: + +H1: Teams using the framework will show higher TTI scores during the intervention period than during the baseline period. + +H2: Teams using the framework will detect communication-to-record inconsistencies faster than during the baseline period. + +H3: The framework will increase documentation completeness without significantly increasing perceived cognitive burden. + +H4: Role-aware confirmation and provenance links will improve perceived trust in AI-generated summaries and updates. + +### 5.5 Data Analysis + +Quantitative analysis should compare baseline and intervention periods using within-team changes and, where available, between-team comparisons. Because team sample sizes may be small in early studies, effect sizes and confidence intervals should be reported alongside significance tests. Qualitative data should be coded for themes such as trust, interruption cost, perceived usefulness, correction behavior, and concerns about surveillance or accountability. + +## 6. Ethical, Privacy, and Governance Considerations + +Conversational AI in team communication creates governance risks. A system that records meetings and chat can easily shift from support to surveillance if boundaries are unclear. The framework therefore requires explicit policies for consent, data retention, role-based access, and auditability. + +Four safeguards are essential. First, team members should know which channels are monitored and what data is stored. Second, sensitive personal content should be redacted or excluded when it is not relevant to project coordination. Third, AI-generated summaries should retain source links and confidence levels. Fourth, team members should be able to correct, reject, or reverse AI-proposed records. + +The system should also avoid treating transparency as individual performance scoring. The TTI is designed as a team-level process metric. Using it to rank individuals would distort communication behavior and could reduce psychological safety. + +## 7. Discussion + +The framework reframes Agile transparency as a traceability and consensus problem. Rather than asking teams to produce more documentation manually, it uses AI to identify where documentation already exists implicitly in conversation and where that knowledge conflicts with official records. This approach may reduce documentation gaps, but only if the system is designed with interruption discipline and human control. + +The framework also clarifies the boundary between AI assistance and AI decision-making. The AI system extracts, links, and proposes. The team confirms, rejects, or revises. This boundary is important because Agile decisions often depend on tacit context, stakeholder priorities, and team norms that cannot be inferred reliably from text alone. + +The proposed TTI provides a starting point for measurement, but it should be interpreted carefully. High coverage is not always good if the system captures trivial or irrelevant items. Fast writeback is not always good if updates are pushed before agreement. The index is most useful when combined with qualitative feedback and governance checks. + +## 8. Limitations and Future Work + +This paper is limited to a conceptual framework and evaluation protocol. It does not report deployment results, statistical outcomes, or validated performance gains. The original version of the manuscript included empirical-sounding language and expected percentage improvements; those claims have been removed or reframed as hypotheses because no supporting data were provided. + +Future work should implement a prototype, run a pilot study, validate the TTI components, and evaluate whether the system improves alignment without increasing interruption burden. Additional research should examine cross-team deployment, integration with different project management systems, multilingual teams, and the effect of AI mediation on psychological safety. + +## 9. Conclusion + +Agile teams already generate rich knowledge through meetings, chat, issue updates, commits, and reviews. The problem is that this knowledge is fragmented and often unverifiable after the fact. This paper proposes a conversational AI framework that transforms informal communication into validated, traceable, and team-approved knowledge artifacts. By combining NLP extraction, artifact linking, conflict detection, and role-aware confirmation, the framework supports team memory without replacing human judgment. The proposed Team Transparency Index offers an initial measurement model, while the evaluation protocol defines a path for empirical validation. + +## References + +Cinkusz, K., Chudziak, J. A., & Niewiadomska-Szynkiewicz, E. (2025). Cognitive agents powered by large language models for agile software project management. *Electronics, 14*(1), Article 87. https://doi.org/10.3390/electronics14010087 + +Itzik, D., & Roy, G. (2023). Does agile methodology fit all characteristics of software projects? Review and analysis. *Empirical Software Engineering, 28*, Article 105. https://doi.org/10.1007/s10664-023-10334-7 + +Malla, P. (2025). Analyzing the impact of agile methodologies on software quality and delivery speed: A comparative study. *World Journal of Advanced Research and Reviews, 25*(1), 1207-1216. https://doi.org/10.30574/wjarr.2025.25.1.0184 + +Stray, V., Moe, N. B., & Bergersen, G. R. (2017). Are daily stand-up meetings valuable? A survey of developers in software teams. In H. Baumeister, H. Lichter, & M. Riebisch (Eds.), *Agile Processes in Software Engineering and Extreme Programming* (pp. 274-281). Springer. https://doi.org/10.1007/978-3-319-57633-6_20 + +Stray, V., Moe, N. B., & Sjoberg, D. I. K. (2020). Daily stand-up meetings: Start breaking the rules. *IEEE Software, 37*(3), 70-77. https://doi.org/10.1109/MS.2018.2875988 + +Stray, V., Sjoberg, D. I. K., & Dyba, T. (2016). The daily stand-up meeting: A grounded theory study. *Journal of Systems and Software, 114*, 101-124. https://doi.org/10.1016/j.jss.2016.01.004 + +Umar, M. A. M. A., Lano, K., & Abubakar, A. K. (2025). Automated requirements engineering framework for agile model-driven development. *Frontiers in Computer Science, 7*, Article 1537100. https://doi.org/10.3389/fcomp.2025.1537100 + +Verwijs, C., & Russo, D. (2024). Do Agile scaling approaches make a difference? An empirical comparison of team effectiveness across popular scaling approaches. *Empirical Software Engineering, 29*, Article 75. https://doi.org/10.1007/s10664-024-10481-5 + +## Revision Log + +| # | Integrity issue | Action taken | Status | +| --- | --- | --- | --- | +| 1 | Abstract claimed a completed mixed-method evaluation without evidence. | Rewrote abstract to identify the manuscript as a conceptual framework and evaluation protocol. | Resolved | +| 2 | Unsupported percentage gains appeared as expected impact. | Removed performance claims and converted evaluation expectations into testable hypotheses. | Resolved | +| 3 | No formal References section. | Added APA-style references with DOIs or publisher links where available. | Resolved | +| 4 | Framework lacked a clear empirical validation plan. | Added evaluation protocol with study design, measures, hypotheses, and analysis plan. | Resolved | +| 5 | AI role risk was underdeveloped. | Added governance, consent, reversibility, and anti-surveillance safeguards. | Resolved |